Turn?» ..._ ..__.__ LIB R A R Y Michigan State University This is to certify that the thesis entitled Markov Random Field Texture Models presented by George Robert Cross has been accepted towards fulfillment of the requirements for P’\' ‘ D ' C({M’L'va’lf 5(1)! GWQ ‘1’ degree in M \IquQ-"l (7?be Major professor Date H/ iL/S’C’ 0-7639 S w: 25¢ per day per item RETURNING LIBRARY MATERIALS: Place in book return to remove charge from circulation records MARKOV RANDOM FIELD TEXTURE MODELS BY George Robert Cross A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Computer Science 1980 ABSTRACT MARKOV RANDOM FIELD TEXTURE MODELS BY George Robert Cross We consider a texture to be a stochastic, possibly periodic, two-dimensional image field. A texture model is a mathematical procedure capable of producing and describing a textured image. A detailed account of the current literature about texture models is given, along with a discussion of the intuitive notions of texture. We explore the use of Markov Random Fields as texture models. The binomial model, where each point in the texture has a binomial distribution with parameter 9 controlled by its neighbors and "number of tries" equal to the number of gray levels, was taken to be the basic model for the analysis. This represents the first use of the binomial model for image analysis and the first application of the binomial model for any purpose. A method of generating samples from the binomial model is given followed by a theoretical and practical analysis of its convergence. Examples show how the parameters of the Markov Random Field control the strength and direction of the clustering in the image. The power of the binomial model to produce blurry. sharp, line—like. and blob-like images is demonstrated. Natural texture samples were digitized and their parameters were estimated under the Markov Random Field model. The use of a hypothesis test' for an objective assessment of goodness-of-fit under the Markov Random Field model is a key feature of this investigation. Overall, highly random textures fit the model well. The estimated parameters of the natural textures were used as input to the generation procedure. The synthetic micro-textures closely resemble their real counterparts, while the regular and inhomogeneous textures do not. Co-occurrence matrices of binary, first-order Markov Random Field textures were computed as a function of the field parameters. There is a good correspondence between the predicted matrices and the observed matrices on numerous samples. Ala realistic_appraisal of the Markov “f as a texture model is accompanied by a for future study. Random list of To Ellen ii ACKNOWLEDGEMENTS I thank my major professor Anil K. Jain for his help and encouragement during the preparation of this thesis andr more important, for guiding me toward a career in Computer Science. Professor Richard Dubes deserves special recognition for his careful reading of numerous preliminary drafts and suggestions for improvement. I also thank Professors Page and Mandrekar for serving on my committee and Professor Hedges for welcoming me into the department. Thanks are due in addition to the PR/IP Laboratory graduate students for their assistance during the hectic days of thesis preparation: Neal Wyseo Steve Smith, Wei-Chung Lin, and Gautam Biswas. The partial support of NSF Grant ECS-8007106 is also gratefully acknowledged. iii TABLE OF CONTENTS LIST OF TABLES ................................... vii LIST OF FIGURES ..3............................... viii LIST OF SYMBOLS .................................. x CHAPTER 1. INTRODUCTION ......................... 1 1.1 Intuitive Notions of Texture ............. 2 1.2 Texture Formation Paradigms .............. 6 1.3 Vision Research About Texture ............ 8 1.3.1 Julesz' Work on Texture .............. 9 1.3.2 Marr's Work on Texture Vision ........ 13 1.4 Applications of Texture .................. 14 1.5 Organization Of the ThESTS ooooooooooooooo 20 CHAPTER 2 TEXTURE MODELS ...............5......... 21 2.1 Introduction ............................. 21 2.2 Stochastic Texture Modelling ............. 26 2.2.1 Ouantized Continuous Models .......... 27 2.2.2 Time Series Models ................... 29 2.2.3 Fractal Textures ..................... 30 2.2.4 Random Mosaic Models ................. 32 20205 Syntactic MethOdS oooooooooooooooooooo 37 iv 2. 2. 2.3 CHAPTER 3.5 3. 3. 3. 3. 3.6 CHAPTER 4.1 4.2 4.3 CHAPTER 2.6 Linear MOdELS oooooooooooooooooooooooo 207 Markov MOdElS oooooooooooooooooooooooo Summary oooooooooooooooooooooooooooooooooo 3. MARKOV RANDOM FIELDS ................. Introduction ............................. Definitions and Theorems ................. Further Assumptions ...................... Order Dependence ......................... Specific Models .......................... 5.1 Binary Model ......................... 5.2 Binomial Model ....................... 5.3 Other Auto-Models .................... 5.4 Other Lattice Models ................. 3.5.4.1 Strauss Model .................... 3.5.4.2 Helberry-Galbraith Fields ........ Summary oooooooooooooooooooooooooooooooooo 4. SIMULATION of MARKOV RANDOM FIELDS ... Introduction ............................. Definitions .............................. Simulation Procedure ..................... Convergence Properties ................... Examples of Textures ..................... Summary nooooooooooooooooooooooooooooooooo 5. MODELLING NATURAL TEXTURES ........... 38 4O 42 44 44 45 49 51 59 61 61 62 62 65 66 67 67 7O 78 5.1 5.2 5.3 Introduction ooooooooooooooooooooooooooooo Estimation Of Parameters ooooooooooooooooo HypotheSTS Testing ooooooooooooooooooooooo 504 DeSCertTon Of the Data oooooooooooooooooo 5.5 Evaluation Of the Fit oooooooooooooooooooo 5.5.1 Binary Texture Results ooooooooooooooo 5. 5.6 5.2 Binomial Texture ReSULtS ooooooooooooo Texture Matching Experiments ............. 50601 Binary Textures oooooooooooooooooooooo 50602 Binomial Textures oooooooooooooooooooo 5.7 Summary oooooooooooooooooooooooooooooooooo CHAPTER 6.1 6.2 6.3 6.4 6.5 CHAPTER 7.1 7.2 6. CO-OCCURRENCE MATRICES ............... Introduction ............................. Joint Probability Formulation ............ Join Count Statistics .................... Experimental Results ..................... summary oooooooooooooooooooooooooooooooooo 7. CONCLUSIONS AND DISCUSSION ........... summary oooooooooooooooooooooooooooooooooo DISCUSSION ooooooooooooooooooooooooooooooo 70201 Advantages ooooooooooooooooooooooooooo 7.2.2 Disadvantages and Limitations ........ 7.3 Suggestions for Future Research .......... LIST OF REFERENCES 0.0.0.0.........COOCOOOCOIOOOOO vi 93 100 104 107 107 121 126 126 135 141 142 142 144 147 155 160 161 161 164 167 169 173 10. 11. 12. 14. 15. LIST OF TABLES Textures Used in the StUdy ooooooooooooooooooo Isotropic. First-Order. Binary Tests ......... Anisotropic. First-Order. Binary Tests ....... Isotropic. First-Order. Binary Tests on Second-Order SpaCTnQ oooooooooooooooooooooo Anisotropic. First-Order. Binary Tests on Second-Order Spacing oooooooooooooooooooooo First-Order (Isotropic). Second-Order (ISOtrODTC), Binary Tests oooooooooooooooooooo First-Order (Isotropic). Seconqurder (Anisotropic). Binary Tests oooooooooooooooooo First-Order (Anisotropic). Second-Order (Isotropic). Binary Tests oooooooooooooooooooo First-Order (Anisotropic). Second-Order (Anisotropic). Binary Tests oooooooooooooooooo Best-Fitting Binary Texture Models ........... Eight Gray Level. First-Order(Anisotropic). Second-Order(Isotropic) Tests oooooooooooooooo Eight Gray Level. First-Order(Anisotropic). Second-Order(IsotrOpic) Tests oooooooooooooooo Best-Fitting Eight-Gray Level Texture Models . Binary Texture Characteristics ............... Errors in Co-Occurrence Predictions .......... 106 112 113 114 115 116 117 119 120 123 124 125 128 159 10. 11. 12. 13. 14. 15. 16. LIST OF FIGURES Brick Hall ................................... Coarse Gravel ................................ Venetian Blinds .............................. Rag Rug ...................................... Repetitions of a Random Texture .............. Squared Distances to the Point X ............. Orders Retative to the Point X .....-......... First-Order Neighbors ........................ Second-Order Neighbors ....................... Constituents of the S(i.k) ................... Conditional Probabilities for the Welberry-Galbraith Field ooooooooooooooooooooo Algorithm for Generating a Markov Random Field ................................. Convergence of b(1..) ........................ Behavior of the Number of Exchanges .......... Isotropic First-Order Textures ............... Anisotropic Line Textures .................... Ordered Pattern .............................. Diagonal Inhibition Textures ................. Isotropic Inhibition Textures ................ Clustered Texture with 4-gray Levels ......... viii 12 52 54 56 60 65 77 80 81 87 88 88 89 90 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. Clustered Textures with 16 and 32 Gray Levels. ooooooooooooooooooooooo'oooooooooooooo Horizontal Texture with 16 Gray Levels ....... Attraction-Reoulsion Textures with 16 Gray Levels ............................... First-Order Coding Pattern ................... Second-Order Coding Pattern .................. Third- or Fourth-Order Coding Pattern ........ Example of a Chi-Square Test ................. Non-Markovian Periodic Texture ............... Real and Synthetic Binary Textures ........... Real and Synthetic Eight Level Textures ...... Expected and Observed Co-Occurrence Probabilities oooooooooooooooooooooooooooooooo ix 90 91 91 96 97 99 103 110 131 138 158 fines; b(i,.) b(i9k) 881 BB? BN1 SW? C F f(niisj) G Y(1) Y(2) m(j95) LIST OF SYMBOLS Markov Random Field Parameter for {0.1} variables Markov Random Field Parameter for {-1.1} variables Markov Random Field ith-order isotropy parameter Markov Random Field ith-order. kth direction parameter Horizontal black-black join count Vertical black-black join count Horizontal black-white join count Vertical black-white ioin count Co-occurrence matrix Clique function First-passage probability Number of gray levels Markov Random Field horizontal clustering parameter Markov Random Field vertical clustering parameter Lattice or image Mean recurrence times N olivi) {q(j)} my S {S(i)} S(i9k) 5(T) U1 U2 V(i9j) ”WI ””2 N81 W82 5 Xlisj) Z 9 Image dimension Transition probability Coloring probability Conditional probability of X9 given its neighbors Limiting distribution Log probability ratio Markov Random Field mean States of Markov chain Sum of neighbors at order iv direction k Weighted neighborhood sum Binomial model parameter Horizontal join count Vertical ioin count Interaction parameter Horizontal white-white join count Vertical white-white join count Horizontal white-black join count Vertical white-black join count Coloring Gray level at point (i.j) Partition function Coloring of all zeroes CHAPTER 1. INTRODUCTION Image modelling is a central part of an image understanding system. It is also the core of classical image processing: image restoration [6]. and two-dimensional filtering [423. The subject of image modelling involves the construction of models or procedures for the specification of images. These models serve a dual role in that they can describe images that are observed and also can serve to generate synthetic images from the model parameters. We will be concerned with a specific type of image model. the class of texture models. Understanding texture is an essential oart of understanding human vision. In this thesis. we will explore the use of a Markov Random Field model for the generation and analysis of textured images. The goal of the research is to produce a texture analysis and synthesis system which will take as input a stochastic texture. analyze its parameters according to the Markov Random Field model. and then generate a textured image that both resembles the input texture visually and matches it closely from a statistical point of view. This can be considered a kind of Turing test for image generation [1003. in that the proof of the viability of the system will be to produce textures that cannot be distinguished by humans from their real counterparts. We will not perform a rigorous psychological study of the correspondence. but will concentrate on the statistical evaluation of the goodness of fit of the observed texture and the generated texture. 1.1 Intuitive Notions of Texture We mention a study by Tamura et al. [983 which attempted to find statistical features corresponding to the usual attributes of texture. Although the study had limited success. the textural attributes identified serve as a useful framework for the discussion that follows. The study delimited six attributes: coarseness. contrast. directionality. line-likeness. regularity. and roughness. Coarseness refers to the size of the cells (areas of near equal brightness) present in the picture. A fine grain picture has small cells. whereas a coarse picture has large cells. Large areas of equal brightness are significant visual cues. 0‘ The contrast of a picture is determined by local Ivariation in gray tone. A low-contrast picture nay have all the gray scale values present. but has few transitions from near black to near white in object boundaries. Before proceSsing a picture. we often normalize the contrast and modify its histogram so that an equal number of pixels is present at each gray level. Moreover. the gray scale is then adjusted so that G levels occur at equally spaced intervals from 0 (black) to 255 (white). Such a transformation distorts the natural contrast in the image and negates the value of this feature. Directionality refers to whether we perceive the trend of the image to be vertical. horizontal. skewed. or non-directional. The bricks picture. Figure 1. is clearly directional. while the gravel image. figure 2. is non-directional. Line-likeness refers to the presence of lines in the picture. It is contrasted to a blob-like effect. where there are clusters of equal brightness and of circular shape. Some images have a mixture of both. Figure 3 is an image of Venetian blinds and has a definite line-like structure. Eigugg 13 Brick Wall E1935; g; Coarse Gravel Eigggg g; Venetian Blinds £12932 3; pac; Rug Regularity refers to whether the image is highly random or not. An image of a rag rug. Figure 4. is clearly regular. whereas the picture of gravel. Figure 2. is irregular. Finally. roughness describes a tactile attribute of texture rather than a visual one per se. But a rough surface is also visibly rough. and real pictures show surface contours that we would expect to feel rough if we could touch them. 1.2 Texture Formation Paradigms There is no universally accepted definition for texture. Part of the difficulty in giving a definition of texture is the extremely large number of attributes of texture that we would like to subsume under a definition. Moreover. some of these attributes are aooarently contradictory. The first definition is that a texture is a periodic. stochastic gray-scale image. Examples of such textures include cloth. tiles. and bricks. The second definition of texture is any stochastic gray-scale image. This includes such textures as grass. sand. and clouds. Most texture research can be characterized by the underlying assumptions made about the texture formation process. There are two major assumptions. and the choice of the assumption depends primarily on the type of textures to be considered in the study. The first assumption. which is called the placement rule viewpoint. was explained by posenfeld and others [86]. [111]. In this model. a. texture is composed of small primitives. These primitives may be of varying or deterministic shape. such as circles. hexagons. or even dot patterns. The textured image is formed from the primitives by placement rules which specify how the primitives are oriented. both on the image field and with respect~ to each other. Examples of such textures include tilinos of the plane. cellular structures like tissue samples. or the picture of bricks. Figure 1. Notice that no two bricks are identical. but there is a uniformity to their placement. There is a random aspect to the brick picture in that the shading of the individual bricks varies. The image of gravel. Figure 2. is not appropriately described by a placement model. The key feature of this image is the coloring and distribution of the cells (areas of near-equal brightness). The primitives are very random in shape and cannot be easily described. The second viewpoint regarding texture generation processes involves the stochastic assumption. We have already seen that the placement 'rule paradigm for textures may include a random aspect. In the stochastic point of view. however. we take a more extreme position and consider that the texture is a sample from a probability distribution on the image space. The image space is usually an N by N grid and the value at each grid point is a random variable in the range {091100096-1}0 1.3 Vision Research about Texture In this section. we will describe what vision researchers have found to be the important constituents of human texture vision. Although machine perception need not be an emulator of the human visual system. the success of human vision in discriminating textures merits attention. In addition. since we are interested in performing automatically tasks that are currently done by humans. such as radiographic screening. we should be guided by the cues and sensitivities of the human visual system. 1.3.1 Julesz' Work on Texture Julesz' investigations of texture span twenty years. Besides the study of texture perception itself. he has studied texture's role in binocular or stereoptic perception. motion perception. and color perception [573. He will examine only the work on texture itself. Early in his work. dulesz decided that it is appropriate to examine perception of textures without visual cues. He uses random dot patterns controlled by analogs of one-dimensional Markov processes and Gaussian processes. Because of this. his work has limited applicability to textures generated according to placement rule schemes. His work is also concentrated on "pure perception." which means perception in the absence of higher-order scrutiny. When he speaks of "effortless discrimination." he refers to the process of deciding whether two textures are the same after seeing them for only 100 milliseconds. If they are seen for a longer time than that. higher-order conceptual processes take over. One of the key features that he identified in texture perception is cluster formation [56]. This is related to image coarseness in the sense mentioned earlier in section 10 1.1. He found that the human visual system acts as a slicer or thresholding device in clustering intensity values. For example. if the image contains the gray levels 0. 1. 2. and 3. then it is not possible for clusters to form consisting of the gray level sets {0.2} and {1.3}. Further. cluster detection is central to the entire perceptual process. from the lowest level aggregation up to the higher order mental processes. The aspect of dulesz' work that has captured the most attention is the conjecture that the human visual system cannot effortlessly discriminate textures that agree in their second-order statistics [54]. [55]. E583. E59]. [60]. The first-order statistics are simply the proportions of the pixels at each gray level. The proportions control whether we see the picture as a black object on a white background or vice versa. Julesz has found that we can effortlessly perceive a first-order difference between textures within fairly narrow ranges [54]: this work is related to the classical figure-ground problem and numerous optical illusions [35]. The second-order statistics involve the joint probability that a randomly thrown rod of length r will fall on the image in such a way that one of its endpoints lands on a pixel of color i while the other lands on a pixel of color j. 11 Similarily. third-order statistics have to do with triangles and the configuration of gray levels at the vertices after a similar random experiment. The conjecture that the second-order description (which of course implies the first-order statistics) determines our discrimination ability has held up remarkably well with only a few classes of unusual textures that confound it [58]. Another aspect of texture that dulesz investigated [57] is that periodicity of a texture is only observable if the frequency is large compared to the size of the picture. For example. it is all but impossible to notice that Figure 5a consists of the same texture repeated four times. whereas it is immediately obvious that Figure 5b consists of repetitions of the same random structure. The difference is that Figure 5b consists of sixteen repetitions of a 16 by 16 picture in a frame size of 128 by 128. whereas Figure 5a consists of only four repetitions of a 64 by 64 picture in the same frame size of 128 by 128. Julesz' work has been extended by a number of researchers. notabiy Pratt ef al. [83]. It is now generally assumed that- the second-order statistics are 12 yxag‘xYXxfi-ifi; -..-.6 4 w‘ was Wm?) gflfigk‘r A . " g ‘ ‘qy‘fi ‘ 3‘, 4‘ "9'7: i ) Em. « $3}in ”K‘sf“ My I.) «9‘ 1’ «’3 J; :3‘¢;«_“ kg‘Ya‘: ‘w‘y 964433;: a; Figure 5. Repetitions of Random Textures: (a) Four repetitions (b) Sixteen repetitions. sufficient in a pattern recognition environment to discriminate between textures. There is still much work to be done in the sense that the second-order probabilities must be combined in some way to provide features for recognition of the textures. The previously discussed work of Tamura et al. [98] is such an attempt. as is the continuing efforts of Conners and Harlow [23] to define a texture recognition feature set from the second-order statistics. It is possible to use third-order statistics in texture discrimination even though the human visual system does not appear to make use of them. The difficulty lies in attempting to estimate 13 the third-order parameters given a relatively small texture field. such as a field of size 64 by 64. Moreover. the storage requirements for maintaining frequency statistics on the large number of possible configurations are enormous. 1.3.2 Marr's Work on Texture Vision The ultimate goal of Marc's research. is a comprehensive computational theory of vision. Such a theory can be implemented with a computer. which will. in fact. emulate the human visual system. His plan revolves around the "primal sketch" E68]. [693. The primal sketch is produced from the original intensity values received at the retina. The primal sketch consists of measurements on the strengths and characteristics of the following primitives: edges. lines or bars. and blobs. These primitives are quantified further by the predicates of orientation. size (length. width. diameter. or aspect ratio if appropriate). contrast. position. and terminal points. This is in general agreement with the work of Caelli and dulesz E18]. [19] in describing the processing at early levels in the visual system. 14 Marr asserts that once the primal sketch is computed. it can then be analyzed to find objects or regions of interest. This is done by aggregating the primitives which are close in some sense. We may thus characterize his overall approach as "bottom-up." in the sense that we do not begin with high-level knowledge of the scene or preconceptions of the objects present. The recognition of objects using the primal sketch model is based on using the texture information to define the objects. rather than considering the texture to be noisy background. 1.4 Applications of Texture There are essentially four areas of image processing in which texture plays an important role. We offer this discussion as justification of our interest in the production of synthetic textures. As a side benefit. such generation procedures may help in furthering our understanding of human perception. According to Julesz C5739 ... to solve the problem of visual texture generation of familiar surfaces is important for both theoretical and practical uses. I can only hope that in the near future scientists will learn to clarify the enigmatic problem of familiar textures. 15 (i) Classification_ Suppose. for example. that we have a series of pictures of mineral samples that we would like to classify into groups based on the type of mineral represented. The true class of some training samples is generally known and could be provided by geologists. Such images have a special structure characteristic of the mineral. but the structure is not deterministic like a tiling. We would like to specify some features of the image in order to classify it. There are obvious differences in distribution of grain size. gray tone. and even color. Studies of this type have been done by Haralick [46] on sandstone samples and by Weszka et al. [106] on terrain samples. (ii) Image Segmentation Segmenting an image means dividing it into homogeneous regions such that each region has some significance. Typically. the regions are connected and we have some a 951951 idea of what we are looking for or what might be present in the image. For example. we may wish to segment a LANDSAT image into land-use categories by means of texture analysis. Nevatia [76] uses the 16 assumption that the objects of interest have strong boundaries and the textured background is composed of short line segments to extract the boundaries of a small toy tank from the background of grass. Thompson E°93 uses a generalized edge operator to define texture borders and is successsful in finding the boundaries between two regions with differing textures. A linear approach to segmentation of textured regions is taken by Deguchi and Morishita [26]. which we will discuss in detail in chapter 2. (iii) Realism in Computer Graphics Computer graphics differs from image processing in that its goal is the production of images frOm a description. whereas image processing is concerned with the modification and interpretation of real-world images. Although high resolution. color. raster-graphics displays capable of producing realistic images are available. the existing algorithms cannot produce the desired realism. Texture gives important information on the depth and orientation of an object. besides being an intrinsic feature of realistic objects [643. [78]. A number of approaches have been taken. mostly related to mapping a 17 flat textured image onto the skeleton of a three-dimensional object and then displaying a perspective view. Catmull E20]. [21] describes a procedure for wrapping photographs of texture onto objects to produce textured surfaces. Blinn [14]. [15] uses a patch of texture to tesselate the image. Csuri et al. [25] extend the basic patch idea to include the specification of structure for the patch in terms of reflectance values at each pixel. This allows more realistic shading and permits the simulation of the appearance of the textured patch under the generated illumination. They intend to extend their model to include some randomness in the patches. Dungan [293 has coupled height data from Defense Mapping Agency maps with an overlay identifying terrain segments (trees. lakes. etc) to simulate the appearance of the ground scene to a viewer in an airplane. As the applications of computer graphics grow. the importance of an understanding of texture will also increase. In fact. the generation of natural textures will likely play a more important role than the recognition of textures. 18 (iv) Picture Encoding The final area where texture generation and analysis has application is in picture. encoding. If we can generate a texture with a few parameters that is indistinguishable from an observed texture. then we have effectively compressed the enormous amount of data in the original texture to the parameter vector. By analogy. this is tantamount to knowing an algorithm for the generation of normal deviates rather than carrying an enormous pre-printed book listing millions of samples from a normal population. Such a procedure has application to the previously mentioned graphics problems in that we could artificially generate background scenes from descriptions defined by textural parameters. Such scenes are presently generated by a hybrid graphics and image processing procedure of mapping photographs onto the graphics skeletons. Some work has already used texture synthesis in the pure information theory context of reducing the total number of bits needed to store the picture. Modestino and Fries [75] encode a picture according to a poisson line model. while Delp et al. [27] use a Gaussian model. In 19 both cases. the image is partitioned into small squares. The texture of each square is estimated and the few parameters of the texture are stored instead of the 256 gray levels of the observed texture. Since the picture is reconstructable by the same model that was used to estimate it. the picture can be approximately reconstructed square by square from the estimated parameters. 20 1.5 Organization of the Thesis Chapter 2 contains a review of texture models that have appeared in the literature and further specifies the levels of texture modelling. Chapter 3 describes the ’specifics of the Markov Random Field theory needed to understand the models that we use in generating and fitting textures. Chapter 4 details the generation procedure for obtaining a texture sample from a distribution specified by a parameter set. In chapter 5. we give our results in fitting a variety of models to a sample of textures from the Brodatz texture album [17]. Chapter 6 gives theoretical results in deriving vtextural features from the model parameters. such as the co-occurrence matrices for Markov Random Field textures. Finally. chapter 7 gives our conclusions and directions for future research. CHAPTER 2. TEXTURE MODELS 2.1 Introduction By a model of a texture. we mean a mathematical process which creates or describes the textured image. The goal of texture modelling is the description of the image; real textures can be compared to generated textures as a test of the validity or utility of the model. The test can take the form of a psychological study or of a statistical assessment of goodness of fit. A secondary goal of texture modelling is classification of textures. The numerical parameters of the model can be used as features to classify the texture. There is a distinction between model-based studies and attempts to find good features for the classification of textures. In a model-based environment. we have the capability to produce. for example. textures that match observed textures. In a feature-based texture analysis. the textural features are measured without an ideal or representative texture in mind. 21 22 Texture modelling can be approached at three levels. The highest level. which we call} the knowledge representation level. uses our understanding of the physical scene in order to describe the textured image. For example. consider the set of textured images of wood grain observed in samples {T(e}). where 9 is the angle at which the cut was made in the tree to form the wood sample. Because trees grow in concentric rings. the ideal appearance of such samples T(9) varies continuously from an image of concentric circles (when 9 is O) to ellipses. and finally to parallel lines (when 9 is 90 degrees). This ideal image is. of course. corrupted by noise. degraded by irregularities in the circles. and. perhap. distorted from the expected image by the presence of knots. We do have. however. an ideal image for specifying membership in the class defined by 9. Such a model was used in automatic drafting [1083. Once we have this ideal model. the classification problem can be attacked. In the present example. the classification problem has been reduced to the relatively trivial one of circle and line detection along with estimation of the aspect ratio of the ellipses. Thus. our structural description of the expected image permits us to identify observable features. 0n the other hand. the 23 knowledge used in one domain of images may not be applied easily to other areas. Models constructed for the texture of corn fields from satellite photographs may not be of much use in constructing a model for the appearance of wheat fields or even corn fields grown in other countries under different cultivation regimes. The success of the knowledge representation approach is limited by our capability to generalize. The bottom level of texture modelling is called the feature-based level. In fact. it is possible to make a case for not calling feature-based. studies modelling at all. We assume that the textures are samples from some unknown probability distribution over a lattice or grid L. representing the image. which has dimension N by N. The objective is to find features that correctly classify a texture into one of the texture classes represented by a set of different probability distributions. A survey of the well-known and commonly used texture features that have been investigated appears in a recent report [47]. Texture features include the image autocorrelation function [62]. which has been found to have visual significance. An influential paper by Haralick [45] describes the use of features computed from the gray level 24 co-occurrence matrices. The (r.s) entry in a co-occurrence matrix is the probability that a point of the image (i.j) has gray value r while point (k.l) has gray value 5. where k = i + dx. and l = j + dy. The increments dx and dy are constant over the individual matrices and we compute a set of matrices indexed by the increments (dx.dy). If G is the number of distinct gray levels in the image. then the co-occurrence matrices are of size G by G. Haralick [45] defines some 14 features based on the co-occurrence matrices. We mention three of them. Let R be the number of pairs of grid positions that were used to compute the co-occurrence matrix C with entries c(i.j) corresponding to the displacement (dx.dy). The angular second moment feature fl is defined by 6-1 6-1 \ \ 2 \ \ / c(i.i)\ f1 = --------- / / \ r / / / and measures the homogeneity of the image. Large values of f1 indicate the presence of many dominant gray level transitions in the image. The contrast feature is given 25 by f2 G-1 \ 2 \ 2 \ n \ / c(i.j)\ f2 = ----- "" / / \ r / / / n = 0 li-3l=n and is a measure of the local correlation. Small values of f2 correspond to images with little local variation. Finally. the correlation feature f3 measures the linear dependence in the image: 6-1 6-1 \ \ t \ \ f3 = (ijc(i.j)/R) - Uny / / ------------------- / / SxSy i = 0 i = 0 where Ux and Uy are the means and Sx and Sy are the standard deviations of the marginal distributions associated with c(i.j)/R. The feature f3 is large if the correlation in the direction expressed by (dx.dy) is large. 26 Several studies have shown that features based on co-occurrence matrices perform reasonably well in classifying some textures. E106]. E231. Conners and Harlow [241 have theoretically evaluated co-occurrence features in distinguishing Markov scan textures. Other features that have been used with some success include gray level run lengths E37]. edge per unit area [861. E87]. extrema in gray scale height [86]. and encoded maxima along scan lines [743. 2.2 Stochastic Texture Modelling We now turn our attention to the middle level of texture modelling. the stochastic process approach. We use the term stochastic process in a loose sense to describe a random procedure used to generate an image. There is some inevitable overlap between the stochastic process approach and the knowledge representation point of view since we try to find a stochastic process for the modelling of the image that is physically meaningful and related to the textures which we are modelling. We use more prior information about the texture in the stochastic process approach than in the feature-based approach. In the stochastic process approach. brightness levels or pixel gray values are the random variables. The level 27 X(i.j) at some point (i.j) is not independent of the levels at other points in the image. In fact our principal concern is about the correlations between the {X(i.j)}. This section is subdivided into discussions about the major classes of texture models that have appeared in the literature. A short description of each type is given. along with some indication of the class of textures that can be handled by the model. 2.2.1 Quantized Continuous Models Suppose that we have M points in the unit square. We impose an N by N grid on the unit square and then count the number of points that fall in each grid square. Such a grid can now be quantized by a scale change to form an image in the usual sense. with a gray scale in the range from 0 to 6-1. The value of G can be either the maximum number of points that fall into a grid square or we may choose G a pgiggi and then use a quantizing function to map the accumulated grid counts to levels from 0 to G-1. 28 A wide variety of continuous_processes. including the multivariate normal distribution. are available to form such images. 0n the other hand. much greater generality can be achieved by the use of spatial processes as explained by Ripley £841 or Matern [701. Following Ripley [84]. a realization of a spatial process is a countable set of points without limit points. For each measurable set A. let Z(A) be the number of points of a realization that fall in the set A. If A is a bounded set. then Z(A) is finite. The process is described by the set of random variables (Z(A)! A is a bounded Borel set}. If the process is second order. translation invariant. and rotation invariant. then Ripley calls the process a 'model'. These latter assumptions are for mathematical tractability. Ripley remarks that the first and second moments of a spatial model should be sufficient to distinguish realizations and justifies this by the work of Julesz [54]. who found that. for the most part. textures with the same second-order statistics were indistinguishable by human vision (see chapter 1). The continuous models of Ripley [84] can be broadly delimited on a scale of clustered. random. and inhibitory. Let U be a set of unit area. The intensity of the process 29 is called x and we assUme that EEZ(U)J = x . For a given A. we would expect a concentration of nearest neighbor distances for a clustered process at levels far below that for a random (i.e. Poisson) process. Similarly. an inhibitory model has a concentration of nearest neighbor distances far above that expected for the random process. One kind of inhibitory model is called hard core. in that we do not allow any points less than a. distance 2r apart. where r is a positive parameter of the process called the inhibition distance. The discretization of continuous processes as image models of texture has not been explored except for a brief mention in [90] and some simulation in [11]. Some results in discrete analogs of continuous processes appear in Rogers' monograph on retail trade [85]. but no direct application to images was made. 2.2.2 Time Series Models McCormick and Jayaramamurthy [66] have developed a model of texture based on a time series forecasting technique. A time series is. of course. a one-dimensional sequence of random variables indexed by a parameter which is usually associated with time. To make such a sequence 30 into an N by N image. the time series is broken up into pieces of size N and then stacked as rows of an image. McCormick and dayaramamurthy use the auto-regressive integrated moving average process (ARIMA) to simulate textures. The parameters are estimated from real textures. They exhibited some success in generating textures characterized by long streaky lines such as the cheesecloth picture in Brodatz [17]. image 0105. The applicability of this model to isotropic textures. i.e.. ones with random or irregular globular clusters. is dubious. It is. however. an appropriate model for line-like textures. as they are able to satisfy the criterion expressed in Chapter 1 that the texture synthesis process be successful in the production of an image that the viewer might think was a real texture. 2.2.3 Fractal Textures The term fractal was coined by Mandelbrot [67] to describe point sets or stochastic processes whose Hausdorff dimension exceeds their topological dimension. The topological dimension of a point set usually corresponds to our intuitive notion of dimension but 31 differs for certain pathological cases. As we should expect. the topological dimension of a countable set of points is 0. of a line is 1. and of a plane is 2. A discussion of the definition and the development of the theory of topological dimension is given by Hurewicz and Wallman [51]; a more modern treatment with considerable detail on special cases of dimension in abstract spaces is given by Pears C81]. The Hausdorff dimension of a set is always at least as large as its topological dimension. Sets that are extremely irregular. such as nowhere-differentiable curves and surfaces. have Hausdorff dimensions which exceed their topological dimensions. For example. the Cantor set has Hausdorff dimension equal to lOGZ/log3. though its topological dimension is- 0. Moreover. the Hausdorff dimension of a set in Euclidean D-space can be a fractional value but is always less than or equal to D [51]. A fractal set process can generate a texture in a number of ways. Mandelbrot [671 exhibits some Brownian textures. These are simulations of mountain ranges Z=f(X.Y) over the X-Y plane such that every plane cut in a direction perpendicular to the X-Y plane gives an ordinary planar Brownian function. The displayed image is produced by simulating a Light source at an angle of 60 degrees 32 with the resultant shadow effects. and viewing angle of 25 degrees. The pictures look like irregular surfaces. such as lunar lanscapes. Other textured images can be obtained by dropping on a plane circles whose radii vary according to a random variable with a Pareto distribution. with the exponent in the density function playing the role of the Hausdorff dimension. Work is needed in the estimation of the fit between natural textures and fractal textures before fractal models can be utilized as true texture models rather than just pattern generation processes. In_ addition. further work is needed on the estimation of the Hausdorff dimension of stochastic processes. For the appropriate background see Billingsley C12]. [13]. 2.2.4 Random Mosaic Models A tiling or tesselation of the plane is a collection {A} of disjoint sets whose union is the entire plane. The elements of {A} are called cells. A mosaic is a tesselation combined with a function H from {A} to the finite set of gray levels {0.1.....G-1} which is constant on each cell: H is called the coloring function. 33 We may speak of random tesselations where the tesselations are determined by some random process or where the H function is probabilistic. for example controlled by a discrete Markov process. We generally only deal with mosaics on bounded sets; such mosaics can be achieved by taking a full planar mosaic and intersecting C with the bounded set and restricting the coloring function. Such a restriction is the true environment for the image processing applications. yet it is usually easier to deal with the theory in the infinite plane case and hope that the bounded set lattice is large enough to allow reasonable approximation. In a sense. the mosaic models generalize the standard notion of a digitized image. The pixels of an image correspond to the cells of a mosaic. but the mosaic models allow us to consider cells that are much larger and of different shape. Ahuja [2] makes a distinction between region-based models and pixel-based models. The appeal of the region-based models is that they have the virtues of the stochastic approach along with some of the strength of the placement paradigm in describing large primitives. For example. the bricks picture. Figure 1. is an example of a macro-texture appropriately described by a region-based model. The relatively fine-grained gravel 34 picture. Figure 2. is a micro-texture. best suited to a pixel-based model. The cells can be generated in a number of ways. We follow current literature [75]. E89]. E90] and discuss the possible patterns. (i) Random Checkerboard: A checkerboard is oriented at an angle 6 to the x-axis. (ii) Occupancy Model: Use any ‘spatial ,point process to generate a realization in the plane. Define the cells V(p). indexed by the points p of the realization by the relation: V(p) = {g on the planel d(q.p) minimal among b). Each point of the plane is a member of exactly one V(p) since the realizations have no limit points. This model is also called the Voronoi polygon model. (iii) Poisson Lines: Distribute lines isotropically over the plane using a Poisson process to generate the radii (distance from line to origin) and the angle the line makes with the x-axis. The polygons determined by the lines are convex and constitute the cells of the mosaic. 35 (iv) Bombing Patterns: Let B be a convex set (this restriction is not necessary but makes the process mathematically tractable). The set B is placed on the plane with random orientation at a point x which is determined by a "marked" Poisson process. Marking a process means that we enumerate each point in a realization and assign it a positive integer. We assume that the copies of the bomb B are of one color. say black. while the plane is white. The cells of the mosaic are then the components of the bombing process. After the bombing is over. the plane can be recolored using any coloring function for the components. (v) dohnson-Mehl: Points are dropped on the plane using a marked process and the cells are formed by growth from these "seed" points. Unlike the occupancy model. the resulting cells need not be convex. This model is appropriate as a physical model for crystal growth in metallurgical applications. but has proven to be mathematically intractable for expressing auto-correlations and other features. (vi) Dead Leaves: The dead leaves model is similar to the bombing model except that the boundary of the bombing shape is retained and the boundaries of previously dropped 36 "leaves" are removed. The final pattern resembles the appearance of leaves on the forest floor with many full outlines of the shapes visible and many partial shapes. This model is discussed by Serra C93]. The mosaic models discussed above can be subsumed into a more general theory of random sets called Mathematical Morphology. The theory is under continuing development by a number of European workers and has seen considerable application in the metallurgical area. See the work of Serra [92]. Matheron E71]. E72]. and Giger E403. Serra calls this general class of models Boolean models. The Boolean model starts with a realization of a Poisson process and takes each point to be a growth center. The primary grain is a random set. usually convex. the randomness being controlled by a growth process. It is clear that this very general growth process can include the above mosaic models. Algebraic operations. including the familiar operations of union. intersection and specialized operations called dilation and erosion. allow the specification of composite models. Special hardware is available to perform the set Operations along with the estimation of model parameters [63]. C91]. C95]. [96]. The advantage of some of the mosaic models is that the theoretical autocorrelations can be computed from a knowledge of the process parameters. By fitting the observed autocorrelation to the theoretical autocorrelation for the proposed model. features such as cell size distribution can be inferred without having to segment the image. Julesz [543 has found that the distribution of plateau regions of constant gray level is a very important consideration in texture discrimination. so any model that effectively models cell size contributes to our understanding of texture. 2.2.5 Syntactic Methods The syntactic approach to texture. like the stochastic methods. can be dichotomized into modelling and classification efforts. Ehrich and Foith E303. bolstered by the success of syntactic methods in waveform analysis. have used syntactic methods to encode textures as mountain ranges whose heights are associated with gray levels. The encoded texture can then be used in classification experiments. 38 Our notion of a model includes the capability to synthesize samples from a textural class. Lu and Pu [65]. take this approach and encode 9 by 9 squares of a texture as a primitive in a stochastic grammar and then link the squares together to form an image. Some success is reported in generating Brodatz E17] textures that have a fairly regular structure. such as the picture of reptile skin. We suspect that difficulty would be experienced in using this method to synthesize micro-textures such as sand or grass. Other problems include the extension to multiple gray levels. but the principal problem is the estimation (or in this context. grammatical inference) of the model parameters. 2.2.6 Linear Models Following a procedure analagous to a linear filtering model. Deguchi and Morishita E26] assume a texture model of the form: X('iyj) = A(D!G)X("D.1'O) 4' 5(191) (9.0) (not (090)) where X(i.j) is the gray level at point (i.j). and the 39 E(i.j) are i.i.d Gaussian variables with zero mean. The local neighborhood is a rectangle of dimension 2m+1 by 2n+1. The A(p.q) parameters are estimated by minimizing the expected difference between the estimated X(i.j) and the observed X(i.j). This approach is fairly similar to the classical analysis of Whittle [107]. but Whittle was only interested in estimating the parameters in an applied statistics context' for testing the hypothesis of no interaction between the site variables. When we begin an estimation problem. we do not know what value of m and n to choose. Hence. we must consider them parameters of the model which must be estimated. Deguchi and Morishita use a procedure due to Akaike [4]. E5] in a time series context to estimate m and n in an optimal manner. Textures following the above model can be generated by a filtering scheme. Generate independent Gaussian variables at each site. Then use the filter coefficients A(p.q) to create a new image by summing over the neighborhoods (it is. of course. possible to use an FFT algorithm to do this faster). The real values at each point of the lattice can then be quantized. 4o Deguchi and Morishita do not generate any textures from the above model. Their purpose is to segment the textured regions of an image. One example is the location of small regions. called "defects". whose texture differs from that of a background texture. To locate the defects. first estimate the A(p.q) parameters for the entire image. Since the defect regions are presumed small. they will not significantly influence the estimation of the A(p.q) parameters. The error between the predicted pixel values using linear estimation and the true value is calculated for each pixel in a 2m by 2n neighborhood. If the error exceeds a preset threshold. then the pixel is presumed to belong to the defect region. Variations on this theme are presented for other segmentation experiments. 2.2.7 Markov Models The term "Markov" is loosely applied to any model where the probability of any particular gray level at a site (i.j) does not depend on pixels beyond a "small" neighborhood of (i.j). We may express this by the relation: 41 p(X(i.j)=k|given all other sites) = p(X(i.j)=k|values at neighbors). The term "neighbors" is defined topologically and can be enforced as any configuration surrounding a given pixel. Following Besag [103. or Hammersley [443. a Markov‘ Random Field obeys the above neighborhood condition along with two other conditions. The first is called positivity and says that any assignment of colors to the lattice (integers in the range {0.1.....G-1}) has non-zero probability. The other condition is called homogeneity and says that the conditional probabilities are unaffected by the actual coordinates of the site: only the neighborhood values matter. The Markov Random Field model has been briefly investigated by Hassner and Sklansky E48]. [493. [50]. Their work was limited to an exposition of the equivalence between the Gibbs field and Markov Random Field expressions for the conditional probability distributions (see Spitzer [94] for the proof) and generation of a few examples of textures. Moreover. they limited their attention to the binary case. 42 An alternative to the use of the full two-dimensional Markov Random Field for images is to traverse the lattice along a scan line and provide a direct analog to the usual Markov chain [1]. Connors and Harlow E24] generated textures according to a simple Markov chain on the rows. which produces streaky line textures but ignores the correlations between pixels in neighboring rows. Haralick and Yokoyama E110] generated essentially one-dimensional textures using scans. but provided some correlations between neighboring rows by considering changes in the features computed from the co-occurrence matrices. 2.3 Summary We have defined a texture model to be a process capable of generating or describing a texture. The types of processes capable of modelling texture were delimited as knowledge-based descriptions. stochastic processes. and feature-based descriptions. We concentrated on the middle level stochastic process models and surveyed the models available. These include discretized continuous models. time-series. fractals. random mosaics. syntactic methods. linear models. and Markov models. Each model was found to be useful for certain classes of textures. It is as inappropriate to assume that one class of generation procedures can generate all textures as it is to assume that the Gaussian distribution can explain all data. CHAPTER 3. MARKOV RANDOM FIELDS 3.1 Introduction The brightness level at a point in an image is highly dependent on the brightness levels of neighboring points unless the image is simply random noise. In this chapter. we explain a precise model of this dependence. called the Markov Random Field. The notion of near-neighbor dependence is all-pervasive in image processing. Focusing directly on this property is a promising approach to the overall problem of micro-textures. The Markov Random Field has had a long history. beginning with Ising's 1925 thesis [53] on ferromagnetism. Although it did not prove to be a realistic model for magnetic domains. it is approximately correct for phase-separated alloys. idealized gases. and some crystals [77]. The model has traditionally been applied to the case of either Gaussian or binary variables on a lattice. Besag's paper [103 allows a natural extension to the case of variables that have integer ranges. either bounded or unbounded. These extensions. coupled with estimation procedures. permit the application of the Markov Random Field to image and texture analysis. 44 45 3.2 Definitions and'Theorems Our exposition follows Besag C101 and Bartlett [7]. As before. let X(i.j) denote the brightness level at a point (i.j) on the N by N lattice L. We simplify the labelling of the X(i.j) to be X(i). i = 1.2.....M where M 1: Let L be a lattice. A c lorigg f L “-6-.— (or a go oging g: L with G Levels) denoted X is a function from the points of L to the set {0.1.....G-1}. The notation Q denotes the function that assigns each point of the lattice to 0. Before defining a special type of probability on the set of all 1. we first give the notion of neighbor. Note that the definition does not imply that the neighbors of a point are necessarily close in terms of distance. Definition 3.2: The point j is said to be a agigflggg of the point i if p(X(i)IX!1).X(2).....X(i-1).X(i+1)....X(M)) depends on X(j). 46 Now we can give the definition of a Markov Random Field. ngini iog 333: A Mggkgg Rand m ielg is a joint probabili y density on the set of all possible colorings K of the lattice L subject to the following conditions: 1.Positivity: p(§) > O for all X- 2.Markovianity: p(X(i)l All points in lattice except i) = p(X(i)l Neighbors of i) 3.Homogeneity: p(X(i)| Neighbors of i) depends only on the configuration of neighbors and is translation invariant (with respect to translates with the same neighborhood configuration). We would like to delimit insofar as possible the kind of probability distributions that represent Markov Random Fields. This is the content of the Hammersley-Clifford theorem. We first need a definition. Definition 3:4: A gliggg is a set of points that consists of either a single point or has the property that each point in the set is a neighbor of every other point in the set. 47 It will prove to be convenient to consider the distribution function in terms of the ratio of p(5) to p(Q). Define the quantity 0(5) by: (3.1) m1) = Ln
0 for all j
\
b. q(j) = 1
3
\
c. q(i)p(i.i) = c(i)
/
i
Under these conditions we have
1.Each state S(i) is positive
2.The set {q(i)} is unique is satisfying
a. QbO’C.
3. lim p(niiyi) = q(j) = 1/m(j.j)
n->°°
conditions
73
95931: See Feller (p393. [313). D
Theorem 4.1 justifies the use of Limiiigg
gigigibgiigg as a description of the set {q(i)}. we
describe a procedure for transforming a given Markov chain
with transition matrix P* = {0*(i’i)} to another chain
with transition matrix P = {p(iqj)} such that the new
chain has limiting distribution {0(1)}.
Ihggggm gig: Consider a symmetric, aperiodic.
irreducible Markov chain with transition matrix P*. Let
{d(j)} be a set of positive numbers with sum 1. Then the
Markov chain with transition matrix D has limit
distribution {0(1)}, where P is defined by:
o*(i.i)q(i)/q(i) if q(i)>q(i)
p(ivi) =
p*(iyi) if q(i)2q(i)
\
p(igi) = p*(i9i) + o*(i,i)(1 - q(i)/q(i))
/
where the ' on the summation means summation over all
indices j with q(j)/q(i) less than 1.
74
95991: See Hammersley [433. D
To apply theorem 4.2 to the problem of generating a
sample from a Markov Random Field. we need to identify the
constituents of the Markov chain. The relevant pieces
are:
gigig sgi: All colorings X of the lattice L.
Maigii 31: p*(i.j) = 1/2. 2 is the total number of
colorings.
igiiii: The limiting distribution value for the state i
should be p(X). where p(.) is the joint probability mass
function over the set of all colorings.
In chapter 6, we will study the joint probability
formulation of a Markov Random Field in detail. The
expression is unwieldy but we do not actually need to use
it. because the actual numbers {q(i)} are never needed in
the calculation of the p(i.j) in theorem 4.2; we need only
examine the Set of ratios {d(j)/q(i)} = {v(i)/0(1)}.
Following Besag [10]. we have
75
l-i
heore
33;: Let 5 and i be two colorings of ithe
Markov Random Field lattice L. Then:
M
p(I) p(X(i)=Y(i)IX(1)9X(2)9...,X(i-1)9Y(i+1)ooo.Y(N))
p(X) p(X(i)=x(i)|X(1),X(2)9...,X(i-1)9Y(i+1)g...Y(N))
1
ggggiz‘See Besag [10], p195. D
In general, we are interested in textures with the
same number of pixels at each gray level. Such textures
are the output of a histogram-flattening procedure such as
the Equal Probability Quantizing algorithm described by
Haralick [453. It is therefore appropriate to limit our
state set to those colorings 3 which have a uniform
histogram. In practice. this is done by starting with an
image that is generated by coloring the point (i.j) with
level k. where k is chosen with equal probability from the
set £09192.....G-1}. The convergence to the limit
distribution is unaffected by the choice of initial
configuration: only the rate at which equilibrium is
reached depends on the choice of the initial
configuration.
Given a state (i.e. colorino ) X! we choose the next
X except that the gray values of
state 1 to be the same as
two randomly selected points are interchanged. In the
76
notation of theorem 4.2. all p*(1.j) are equal. so~we need
only look at q(j)/q(i) = p(1)/p(z) = r.' The algorithm is
diagrammed in Figure 12. This algorithm was used by Flinn
C333 and was invented by Metropolis et al. [733.
Any initial assignment of gray levels to the points
of the image can be made. One possibility is to choose
the next state 1 as 5 with one pixel changed at random.
This allows us to begin with any histogram whatsoever and
reach states {S(i)} with the desired frequency. The other
side of this problem is that the choice of parameters of
the Markov Random Field determines the expected histogram.
This means that if we arrange to make the histogram to be
uniform when. in fact. the parameters do not determine a
Markov Random Field with a uniform histogram. the
procedure will converge to a Markov Random Field with
different parameters than intended. In a practical sense.
this is not a significant problem since we usually want to
generate a texture that resembles a given texture. He can
generate a texture starting with exactly the same
histogram as the given texture and use the measured
parameters from the sample.
77
until STABLE
choose two sites X(1). X(2) with
different gray levels
compute p(z)/p(§) = r
switch X(l) get random number 2
and X(2)
1
V
...
+ ...—..-.— 4y —__.——__ it -—..._.——_. 4....-—_. +>——._... 9—“... 4y
I
switch X(l) I
and X(2) l
I
4l._..—.———._.—-—_-—————.———-——.——-———-——.———-———_—_—.——__._——_ it
4}———————_——.—.——————_ 4..._—_._.—- .0.._—_ .p———.—. .0.
4|._____ + _..____._. 4y _____..._ 4y
1>—--‘—-—-
Eigggg lg Algorithm for generating Markov Random Field
with joint probability function p(z). The coloring I is
obtained from the coloring X by switching the values of
the points X(l) and X(2).
78
4.4 Convergence Properties
Theorem 4.2 guarantees that the application of the
algorithm in Figure 12 will eventually result in a lattice
in a state i with the frequency controlled by 9(1). The
practical question is how long this will take.
We first need to define a time-dimension for the
simulation. Suppose the lattice is N by N and let M=N**2.
We consider M attempted exchanges or switches to
constitute one iiggaiigg. Notice that this ignores
attempted exchanges between pixels of the same color. We
have experimental guidelines for the number of iterations
required to achieve a lattice that matches the input
parameters. In general. it was observed that in 10
iterations or less either the number of changes per
iteration drops to one percent of M or the measured
parameters match the input parameters within about 5
percent. These guideleines define the variable "STABLE"
in Figure 12. On a PDP-ll/34 computer. the time required
for one iteration on a 64 by 64 image was two to three
minutes depending on the number of gray levels.
79
We give an example of the convergence for a specific
case. We want to simulate a first-order Markov Random
Field with binary variables with parameters a = -z. b(1..)
= 1 on a 128 by 128 lattice. The notation follows
equation (3.11) of chapter 3. The estimated parameters
b(1..) 'are shown plotted against number of iterations in
Figure 13. The value of the parameter a is not shown
because. as will be proven later. in the uniform histogram
case. a is -2b(1..). The graph flattens rather quickly
and stays within 0.05 of the intended value of 1.0 for
b(1..). Figure 14 shows the number of changes observed
per 256 attempted exchanges. as the estimates were made
every 1/64 iteration. Thus. although in an iteration of M
attempted exchanges we are still observing nearly M/2
changes. the chain is near equilibrium as the observed
model parameters are within a small tolerance of the
intended parameters.
80
LL:
IMHT
F—o
C3¢
E
0 _
5.]
‘ ‘00
Eigggg lg: Convergence of b(1..).
81
JHHMHHWWWWHP
_ _ _
ootoom ooiowfl oa,ocfl 00.0mm 00 am a: om_
wmozqzu
3
DJ
E. u
Pu
a a . a. m3
Exchanges.
Behavior of Number of
82
4.5 Examples of Textures
We present some examples of textures generated
according to various settings of Markov Random Field
parameters. These images are representative of the kind
of results that can be achieved but are not necessarily
attempts to imitate real textures. They should rather be
considered to be an 'alphabet' of Markov RandOm Field
textures. In chapter 5. we exhibit generated textures
matching observed textures.
(i) Clustering effects
Figure 15 shows a series of 64 by 64 binary textures
with various degrees of clustering. This is an isotropic.
first-order model as represented by equations (3.6) and
(3.7). Figure 15a represents binary 'noise' in the sense
that each pixel has probability .5 of being black and .5
of being white independently of all other pixels. In
Markov Random Field terms. this means a is 0 and b(1.1) =
b(1.2) = 0. The value of b(1.1) is increased from 0 in
Figure 15a to 3.5 in Figure 15h. The increase in
clustering is clearly visible.
83
(ii) Anisotropic Effects
Figure 16 shows extreme anisotropy in first-order and
second-order models on a 64 by 64 lattice. In Figure 16b.
a = -2.0. b(1.1) = 1.93. and b(1.2) = .16. The quantity
b(1.1) controls the horizontal clustering so there is a
large amount of line-likeness in that direction. As
b(1.2) is non-negative. a small amount of vertical
clustering is present. resulting in thickened and noisy
horizontal lines. Contrast this with Figure 16a. which
has parameters b(1.1) = -2. and b(1.2) = 2.1. The value
of b(1.2) causes vertical clustering. similar in intensity
to the horizontal clustering of Figure 16b. The
significant difference is that the negative value of the
parameter b(1.1) forces 'clean' vertical lines with
virtually no horizontal clustering.
The decidedly diagonal effect of Figure 16c results
from the use of a second-order structure. The clustering
in the Nw-SE direction is pronounced since the parameter
in this direction is 1.9 while the parameters in all other
directions are quite small.
84
(iii) Ordered Patterns
Many of the applications of the Ising model involve
studying the checkerboard-like patterns obtained with
negative clustering parameters. In this sense. a perfect
checkerboard is ordered and represents a limit state for a
an alloy or magnet [77]. This is illustrated by Figure
17. which has b(1.1) = -2.25. b(1.2) = -2.16 on a 64 by 64
lattice. The most likely configuration is a black pixel
surrounded by four white pixels or vice versa.
(iv) Attraction-Repulsion Effects
An attraction-repulsion process involves ' having
low-order parameters positive resulting in clustering but
high-order parameters negative in order to inhibit the
growth of clusters. If high-order parameters were also
positive. large clusters would result. whereas negative
high-order parameters yields small clusters.
Figure 18 shows the effect of anisotropic clustering
with inhibition. The first-order parameters of Figure 18a
are approximately 0. which accounts for an immediate
visual impression of randomness. On closer examination.
there are many short horizontal and vertical lines. but
85
very -few diagonal joins. The lack of diagonal joins is a
result of negative second-order diagonal parameters. In
Figure 18b. the first-order clustering parameters b(1.1)
and b(1.2) have been increased. resulting in longer
horizontal and vertical lines.
Figure 19 shows two isotropic attraction-repulsion
textures. These are the result of positive first and
second-order parameters and negative third and
fourth-order parameters. As a consequence of 'the
high-order inhibition. the cluster size is relatively
smaller than one would expect if the third- and
fourth-order parameters were zero.
(v) Multiple Gray Scale Textures
The binary textures above illustrate the essential
features of the visual attributes of a Markov Random Field
but are fundamentally unrealistic as textures. We now
turn our attention to the binomial model. We follow the
notation of section 3.5.
It should be noted that many of these images appear
blurry and out of focus. This effect is not due to the
reproduction process but is intrinsic to the model. If
86
there is no inhibition (via negative high-order
parameters). then the binomial model tends to have smooth
transitions from black to white. The binomial
distribution is unimodal and. as a consequence. values
above and below the mean gray value are highly probable
also. This results in a tapering of the gray scale around
maxima and minima. Such a tapering as one moves away from
black or white points has an effect similar to a
neighborhood averaging or low-pass filter.
Figure '20 shows a 4-gray level picture with
considerable clustering. Figure 20a is isotropic and
first-order. whereas Figure 20b is second-order
anisotropic with diagonal clustering. Figures 213 and 21b
represent typical multiple gray scale pictures with
isotropic fourth-order clustering.
Figure 22 shows a pattern similar to Figure 16. but
with 32 gray levels. The resemblance to wood-grain is
apparent. Figures 23a and 23b show the result of
attraction-repulsion processes with multiple gray levels.
Figure 23a has the appearance of reticulated photographic
film due to strong third and fourth-order inihibition.
The diagonality in Figure 23b is a result of strong
repulsion in some directions and clustering in others.
87
55:51:?
Eigggg ii: Isotropic First Order Textures. The b(1..)
parameters are: _(§_)_ 0.0. La; 0.50. igi 0.75. gg; 1.1.
121 1.26. ii; 1.52. igi 1.79. ihi 3.0. In all cases. the
a parameter is -2b(1..).
88
Eigggg i6: Anisotropic Line Textures. The parameters are
igi a: -0.26. b(1.1) = -2.. b(1.2) = 2.1. b(2.1) = 0.13.
b(2.2) = 0.015. ibi a: -2.04. b(1.1) = 1.93.
b(1.2) = 0.16. b(2.1) = 0.07. b(2.2) = 0.02.
(c) a = -1.9. b(1.1) = -0.1. b(1.2): 0.1 . b(2.1)= 1.9 .
b(292)= '000750
Figure i1: Ordered Pattern. The parameters are a = 5.09.
b(1.1) = -2.25. b(1.2) = -2.16.
am?"
“'I[
iagonal Inhibition Textures. The parameters
.19. b(1.1) = -0.088. b(1.2) = -0.009. b(2.1)
are 2g: 3'-
: -1. b(2.2)
2.05. b(1.2)
Figure 1 : D
‘ 2
-1. (b) a =0.16. b(1.1) = 2.06. b(1.2) =
'2.03. b(2.2) = -2.10.
Eigggg 19: Isotropic Inhibition Textures. The parameter
values EFe: 1a; a :-o.97. b(1..) = 0.94, b(2..) = 0.94.
b(3..) = -o.42, b(4..) : - -o.49. 5g; a = -4.6,
b(1..) 2.52, b(2..) = 2.17. b(3..) = -0.78.
b(qyo) = ”0.850
90
Eigggg gg: Clustered Textures with 4-Gray Levels. The
parameters are (a) a = -2. b(1..) = 1.0. 3g; a = -2.
b(1..) = 1.0, b(291) : 1.0 b(2’2) = 0000
Eigggg 2i: Clustered Textures with 16 and 32 Gray Levels.
The parameters are igi 16 levels. a = -2.0. b(1..) =
b(2..) : b(3..) : b(4..) = 0.05. 1g; 32 levels. a =
-2009 b(1..) : b(290) = b(3..) : b(490) : 0.050
91
Eigggg 22: Horizontal Texture with 16 Gray Levels. The
parameters are a = -2.0. b(1.1) = 0.08. b(1.2) = 1.0.
Eigggg 23: Attraction-repulsion Textures with 16 Gray
Levels. The parameters are igi a = -2. b(1..) = b(2..)
: 029 b(390) = b(4oo) : ”001. b) a = ‘2! b(1,.) 30029
5.-..
b(291) ‘ '202. b(292) = .2. b(391) = '0.059 b(392) = 0.059
b(4.1) = -0.05. b(4.2) = 0.05.
4.6 Summary
We have reviewed the notion of a Markov chain and
seen how to obtain a new Markov chain with desired limit
distribution from a relatively arbitrary chain. This
construction led directly to an algorithm for the
generation of Markov Random Fields with specified
parameters. An example of the convergence of the
procedure was given. along with some guidelines on the
number of' iterations required to obtain a Markov Random
Field with the desired parameters. Finally. we presented
a catalog of textures generated according to various
settings of the model parameters.
CHAPTER 5. MODELLING OF NATURAL TEXTURES
5.1 Introduction
In previous chapters. we have discussed the
probabilistic structure of Markov Random Fields and have
shown how samples from Markov Random Fields can be
generated. One of the principal contributions of this
thesis is the implementation of a statistical measure of
the correspondence between an observed texture and a
texture model. No prior study has performed this kind of
evaluation. All prior studies in texture modelling have
considered a model adequate if its parameters yielded good
classification in pattern recognition experiments or if it
was found to be the best-fitting among a number of models
tested. For example. Deguchi and Morishita E263 determine
the best size for a neighborhood in an auto-regressive
scheme. but do not give any overall guidelines on when the
auto-regressive scheme fits the observed texture.
We first explain the method used to estimate the
Markov Random Field parameters and perform hypothesis
tests. This is followed by the results of testing
textures from the Brodatz texture album [173 for a fit to
various Markov Random Field texture models. The final
93
94
section shows textures generated synthetically in an
effort to match natural textures.
5.2 Estimation of Parameters
Following the notation of equations (3.12) and
(3.13). this section explains the estimation of the
parameter set {b(i.k)} from a textured sample using the
binomial model. The binary model is a special case of the
binomial model and is not handled separately.
The technique used to estimate the parameters is
Maximum Likelihood Estimation. Let p(Xl.) denote the
conditional probability p(X=xl Neighbors of X). where X is
a point of the lattice L. The usual log likelihood is
given by
(5.1) l = ln(p(X|.))
where the summation extends over all points of the
lattice.
95
It is difficult to maximize l as stated in equation
(5.1) since the summands are not independent. Besag [9]
provided a solution to this problem. Instead of forming l
as a sum over all points of the lattice. the lattice is
partitioned into disjoint sets of points called g_gigg§.
Each coding is chosen so that its points are independent.
This can be done by adequately spacing the X points so
that if X(i) and X(i) are two points in a coding. then
X(i) is not a neighbor of X(j) in the Markov Random Field
sense.
The number of codings required depends on the order
of the process. We would like each coding to be as large
as possible because the larger the coding. the more
samples are available to estimate the parameters. A
first-order process requires at least two codings for
estimation purposes. as shown in figure 24. A
second-order process requires spacing so that three by
three neighborhoods do not interfere. This yields four
codings as shown in Figure 25. Third- and fourth-order
processes require separation of three units since they are
based on five by five neighborhoods. as shown in Figure
26. There are nine codings in this case.
96
III:IATIII:IAYIII:IA+IIIiIATIIIIIA+IIIII¢ IIIII¢
_
2 1 2 1 2 1 n 2
.
III¢ Illll¢ Iii... III.. III... I'll." III._.
1 2 1 2 1 2 _ 1
_
IIII¢ III-... III... III.. IIIAT III... IIIt.
2 1 2 1 2 1 2
I'll: IIIAT ill: ill... Iii-vii. I'l'lAv 'liav
1 2 1 2 1 2 1
III-IA. Ill-Ii. I'll: ill... .Illullli. IIIAv III...
2 1 2 1 2 1 2
I:IIL.I:IILTIIIILTIIIIL+IIIIL+IIIiLvliIILT
1 2 1 2 1 2 1
. .
IIIII1+IIIIILvl|iZIAvIII:I1+IIIII.+IIIIIATIIIII¢
. .
2 — 1 2 1 . 2 1 - 2
. . .
ii|+ Iiii. iI'Av iii... iiiav iii... ilixv
_ . .
1 . 2 1 2 . 1 2 . 1
. . .
III... III¢ IIIAT III+ III¢ III+ III¢
. _ .
2 - 1 2 1 . 2 1 - 2
. . _
Iii+ Iiit. Ilia. IIIII+ Illi¢ III.’ III._.
. — .
1 - 2 1 2 _ 1 2 _ 1
. _ _
III... III-IA. III.. III.’ IIIAT III+ IIIAV
are two
There
I
I
I
+
I
I
I
2
1
I I
I 1 I
I I
Pattern.
2
I
I
I
1
Coding
A
v
I
1 l 2
I
g 23 First-Order
gs labelled 1 and 2.
2
+-—-+--—+---+-_—+---+---+—--+---+---+—-_+
_--¢---+---+---+-
£192:
codin
I
I
I
I
I
I
97
IIII :-
'IIII4
IIIILY
IIIIlv
Ill-Ill! t4
III:
I'll. AV Illllll AT
IllzlAvIIIllL
I'll'l AT I'll;
III:IATI:II|A
IIIIIA
IIIA
2
III._
1
IIIA
2
III:IA
-—-+-—-+---+--—+---+
1
IIII|+ IIIIIL
T
T
Y
1
7
Y
T
v
IIIIII¢
IIIII¢
..J
'IIII¢
III ¢
IIIIIA
4
+---+---+---+--
I
I 3
I
2
I
1
2
+-—-+---+_--+---+—--+—--+-—-+-—-+--—+_--
IIIII¢
IIIIiv
I I
I 3 I
I I
+-—-+---+-—-+_-
4
3
4
3
4
4
3
T
IIIII+.IIII.+.IIII.+IIIII
-
2 .
.
'II. +
IIIII¢
III:
III:
I'l'|¢
4
llllll:
1Q
2
I
I
I
+---+---+---+---+_-—+_--+---+---+---¢---
I
I 1
I
4
+---+---+---+---+_-_+---+_--+---+-——+-—-+
There are four
0
n
r
e
t
t
a
P
g
n
.1
.d
0 I
PV4
'
r 3
e
d 9
P 2
O
. 9
d 1
n
0 d
C e
P...
Q.[
e
by
5.3
9;I~
S
8.9
P-n
UT...
Qfic
i_0
F_C
98
I:III¢
IIIII¢
IIIILv
II" 47
III Av
IIIILT
IIIII¢
IIIILT
IIII;V
IIIILT
IIIII+IIIII¢
7
I I I
7 I 8 I I
I I I
I
I
I
9
I
I
I
8
I
I
I
7
9
+---+---+--_+_-—+_-—+---+---+-—-+--_+---
I I
I 7 I 8
I I
1
3
2
III"
III"
I I
I 2 I 3
I I
I
I 1
I
3
¢
¢
¢
¢
A
w—
+---+---+-_-+---+-
IIIILV
IIIILT
I
I
I
. IIIILY
IIII+
IIIII¢
IIIILV
IIIILv
IIIILT
IIIILT
IIII;T
IIIIL+
IIII;
III:IA
III:IA
IIIIIA
IIIIIA
IIIIIIA
IIIIL
IIIIA
IIIIIA
IIII;
IIIII:
T
v
T
v
I I
I 7 I 8
I I
.L
v
+---+---+---+_-
IIIII¢
IIIILV
xv
IIII¢
IIIIIAY
IIIII¢
Ill];
7
I
I
I
+---+---+---+---+—
There
or Fourth-Order Coding Pattern.
5192:: aé Third-
are nine codings LabeLled 1-9.
99
The actual estimation procedure is straightforward. Let
[(i) be the log likelihood for the ith coding?
(5.2) [(i) = ln(D(XI.))
XV
where the summation extends over all points X' in the
coding i. In equation (5.2), p(XI.) depends on the order
of the Markov Random Field whose parameters are being
estimated.
We seek to maximize l(i). This is done by finding
values of the parameters a and {b(i.k)} so that
(5.3) Blli)/Ba = 0 and alli)/bb(jgk) = 0
for j = 1,29,...gr and k : 1.2 where r is the order of the
process. If the process is isotropic at some order i.
then we only consider derivatives with respect to b(jq.)
in equation (5.3).
The system of equations (5.3) is solved numerically
by the usual multivariable extension of Newton's method
[52]. We omit the routine expression 4or the derivatives
100
in equation (5.3) and the second partials that are needed
in Newton's method. The stopping criterion for the
iteration process was fifteen iterations. or a change in
the magnitude of successive parameter estimate vectors of
less than 10E-06. Except for a few cases among the
natural and synthetic texture samples. ten iterations
sufficed to provide a residual of less than 10E-06. This
is fortunate. since each iteration of Newton's method in
the multivariate case effectively requires the inversion
of a matrix of size equal to the number of parameters,
which could be as many as nine for fourth-order
anisotropic textures.
An estimate of the parameter vector is obtained for
each coding. Our final estimate of the parameters is the
average value over all the codings. On a POP-11/34
computer an unoptimized estimation procedure requires
thirty seconds per coding for binary textures and eighty
seconds per coding for eight gray level textures. These
timings are averages for textures of size 64 by 64.
5.3 Hypothesis Testing
Note that we really have only one sample from the
unknown distribution p(X) on the set of colorings of the
101
lattice L. From the conditional probability point of
view. each observed configuration of neighbors and the
value of the center point X is a sample. In this sense.
we have M = N**2/k samples of the conditional density
p(xl.) for the N by N lattice L. where k is the number of
codings. We can then perform a chi-square test of the fit
between the expected frequencies for each center pixel and
the observed frequency. The expected frequencies are
computed using the estimated parameters with an
appropriate reduction in degrees of freedom.
An example of a chi-square test is shown in Figure
27. The null hypothesis is:
H0: The texture is a sample from a Markov Random Field
with the estimated parameter set {a.b(i.k)}
while the alternative hypothesis is simply the negation of
H0. The estimated parameters for the example shown in
Figure 27 are a = -4.26. b(1.1) = 2.70. and b(1.2) = 1.5.
The number of cells is G. the number of gray levels. times
the number of observed neighborhood configurations. The
example is a first-order. anisotropic binary lattice. so
there are 2*9 cells possible and all were observed. The
entries in the table are of the form 'observed(expected)'
and the expected entry is computed using the conditional
102
probability distribution with the estimated parameters.
The degrees of freedom are computed by means of the
formula:
(5.4) df = (G-1)*Nc - E
where G is the number of gray levels. No is the number of
neighborhood configurations (in Figure 27. each
neighborhood configuration is a row). and E is the number
of estimated parameters. In the example of Figure 27.
there are two gray levels. nine neighborhood
corfigurations and three estimated parameters. so we have
(2-1)*9-3 = 6 df. We also use the convention of having at
least one expected observation per cell. Thisresults in
a reduction in the number of cells and degrees of freedom.
Since we are performing a number of tests on the same
data (one on each of k codings). there is a great
likelihood of having the hypothesis of a fit to a Markov
Random Field scheme accepted on some codings and rejected
on others. The results of the tests are not independent.
If they were independent. and we performed k tests at a
level a . then the probability of no rejections would be
(1' d )**I(o
103
s<1.1) 5(1.2) x = o x = 1
I o I--- o I 525( 621.)I 5( 9.)I
I o I 1 I 240( 244.)I 20¢ 1e.)I
I o I 2 I 14( 17.)I 8( 5.)I
I 1 I o I 79( 80.)I 18( 17.)I
I 1 I--- 1 I 93( 91.)I 85( 87.)I
I 1 I--- 2 I 21( 14.>I 52< 59.)I
I 2 I o I 4( 3.3I e< 9.)I
I 2 I--- 1 I 12: 17.)I 244< 239.)I
I 2 I--- 2 I ‘8( 8.)I 512< 512.)I
-r 7
Ficure 21 Example of a Chi-Square Test. This is a binary
64 by 64 texture. A first-order anisotropic scheme was
fitted. The results of coding number 1 are shown above.
Chi-square is 12.30957. on 6 df.
Besag [10] suggests a conservative solution to the
problem of ambiguous results. Suppose the most
significant result we observed over k codings was exactly
at level p. The overall significance of the set of tests
is then taken to be kp. For example. in a first-order
isotropic scheme we obtained two chi-squared values of
7.78 (p=0.10) and 11.3 (p=0.01) on 3 df. The value 11.3
104
has exact significance level of 0.01. Since there are two
codings. we take our overall significance level to be
exactly 0.02 = 2*0.01. and would reject the null
hypothesis at the 5 percent level.
In the tables in section 5.5. we have given the
number of "conservative" rejections at the five percent
level for each scheme along with the number of samples of
each texture that had 091.....k rejections over k codings
at the five percent level. This second tabulation leads
to a "liberal" rejection policy of rejecting a Markov
Random Field fit if any coding is rejected.
5.4 Description of the Data
The study of natural textures was based on twelve
pictures from the Brodatz texture album [17]. The plates
used are given in Table 1. Each plate was photographed on
35mm film to form 24mm by 36mm slides. The slides were
digitized in such a way that the small dimension occupied
slightly over 256 pixels of the 480 by 640 image of
Spatial Data Eyecom system. The 256 by 256 images were
split into sixteen non-overlapping 64 by 64 subimages and
also into four 128 by 128 non-overlapping subimages. The
gray scale of each subimage was reduced from 256 gray
105
levels to both two and eight gray levels using Equal
Probability Quantizing E453.
Table 1: Textures
ref
Name
Brick Hall
Ceiling Tile
Pressed Cork
Calf Fur
Grass Lawn
Handmade Paper
pebbles
Beach Sand
Straw Screening
Water
(1)
Hood Grain
Wood Grain (2)
106
Used in the Study.
r to the Brodatz texture album [173.
312:: flames:
D94
D86
04
D93
D9
031
029
069
D70
The plate numbers
107
5.5 Evaluation of the Fit
The textures were evaluated for their fit to various
models. The purpose of this analysis was twofold. First.
we need to validate that the Markov Random Field model is
generally applicable to 'textured images. The second
objective is to formulate some general guidelines {on how
to choose an appropriate model. in terms of order and the
degree of anisotropy and isotropy. to generate specific
types of textures.
5.5.1 Binary Texture Results
Except for the screen texture. some first- or
second-order model was able to give at least ten
acceptances. under the previously mentioned conservative
decision rule. for each texture sample. Detailed analysis
of the screen texture samples showed very few distinct
neighborhood configurations. This is a consequence of its
regularity. The few neighborhood configurations dominate
the computation of the Markov Random Field parameters.
However. the low frequency neighbor probabilities are not
properly controlled by these parameters. resulting in
large chi-square values. Essentially. the histogram is so
skewed toward these configurations that the positivity
‘108
condition of section 3.2 is nearly violated.
As an example of this. Figure 28 consists of 64
repetitions of a random sixteen by sixteen pattern. It
was analyzed on the basis of first- and second-order
isotropic and anisotropic models. Because' of the
repetitions. there are very few different configurations
of neighbors. ”nly 42 different ones appear out of a
possible 81 in a second-order anisotropic scheme.
Chi-squared values are in the range 1000-3000 on 20-40
degrees of freedom. depending on the model. The estimated
parameters vary wildly from coding to coding. For
example. on a second-order anisotropic scheme. the
parameter a was estimated as 0.02. 0.8. -0.243. 0.325.
The b(1.1) estimates were 0.376. -0.469. +0.469. and
-0.567. Similar variability was present in all the
parameters.
Tables 2 through 9 give the results of testing binary
texture samples under a number of models. A test against
a first-order model results in two hypothesis tests. one
for each coding. There are sixteen subimages of each
texture and each may be rejected on zero. one. or two
codings at the five percent level. The number of
subimages which had each number of rejections is recorded
109
in the last three columns of tables 2 and 3. Also listed
is the number of conservative acceptances. of which there
are a maximum of sixteen possible. For an appreciation of
the difference between the conservative and the liberal
rule. compare the number of conservative acceptances with
the number of subimages which had no rejections. Tables 4
through 9 parallel the format of tables 2 and 3 except
that they represent tests performed using the second-order
codings. There is a possibiity of zero. one. two. three.
or four rejections per subimage. as there are four
codings.
Besag [10] makes the point that one cannot compare
the fit of a second-order model to the fit of a
first-order model unless the same coding scheme is used in
both cases. From Tables 2 and 3. it would appear that the
first order scheme fits very well. but in many cases it is
inferior to a second-order scheme when considered on the
same coding. Table 10 gives the best results on
second-order coding for each of the binary textures. The
general good fit of the first order model should be
reconciled with the fact that there are only two codings
rather than the four for a second-order scheme. which
means that fewer rejections are likely.
110
ifiifiit’?
ngHImI
\_I-r 1‘1- 'filv "Iv:
4- -. . '-. 41
Figure 28: Non-Markovian Periodic Texture. This is a 64
by 64 section of a 128 by 128 texture. The basic pattern
is 16 by 16 and is repeated 64 times over the image.
We have limited our attention to the case of 64 by 64
textures. Although third-order estimates can be made
which give good visual results in texture generation
experiments explained later. we cannot reliably perform a
chi-square test of them on the 64 by 64 lattice. The
number of cells with only one member is very large since
the number of possible cells with a fully anisotropic
third-order model is 729 while there are only about 585 in
a single third-order coding (there are 9 codings in all).
111
Our preferred choices for the best-fitting model are
based on a simple rule: choose the model that gives the
We would not consider a fit to be adequate unless the
majority of samples from the texture class fit the model.
112
Table 2: Isotropic. First-Order. Binary Tests. The
parameters estimated are a. b(1..).
Texture Name Conservative Rejections at 0.05 level
Acceptances g 1 2 _
Brick 14 10 6 0
Ceiling Tile 10 9 4 3
Cork 14 12 3 1
Fur 5 2 11 3
Grass 15 12 3 1
Paper 16 12 4 0
Pebbles 14 10 6 0
Sand 15 14 1 1
Screen 0 0 2 14
Water 10 9 7 0
Wood (1) 9 8 8 0
Wood (2) 9 6 9 1
113
Table 3: Anisotropic. First-Order. Binary Tests. The
parameters estimated are a. b(1.1). b(1.2).
Texture Name Conservative Rejections at 0.05 level
Acceptances
Brick ' 11 11 5 0
Ceiling Tile » 11 6 7 3
Cork 11 10 5 1
Fur 6 5 8 3
Grass 13 11 4 1
Paper 13 13 3 0
Pebbles 11 9 5 2
Sand 15 12 3 1
Screen 1 0 5 11
Water 8 8 7 1
Hood (1) 9 7 4 5
Hood (2) 4 2 9 5
114
Table 3: Isotropic. First-Order. Binary Tests. on
second-order spacing. The parameters are a. b(1..).
Texture Name Conservative Rejections at 0.05 level
Acceptances
Brick 16 13 3 0 0 0
Ceiling Tile 16 11 4 1 0 0
Cork . 14 12 3 1 0 0
Fur 11 8 3 5 0 0
Grass 15 11 4 1 0 0
Paper 15 14 2 0 0 0
Pebbles 14 10 6 0 0 0
Sand 14 12 4 0 0 0
Screen 0 0 1 1 5 9
Water 11 10 2 3 1 0
Wood (1) 13 10 6 0 0 0
Hood (2) 13 6 9 1 0 0
115
Table 5: Anisotropic. First-Order. Binary Tests. on
second-order spacing. The parameters are a. b(1.1).
b(1.2).
Texture Name Conservative Rejections at 0.05 level
Acceptances Q 1 2 3 5
Brick 9 6 8 2 0 0
Ceiling Tile 15 11 4 1 0 0
Cork 12 10 4 2 0 0
Fur 10 9 3 3 1 0
Grass 16 11 4 1 0 0'
Paper 14 10 6 0 0 0
Pebbles 11 9 1 6 0 0
Sand 14 10 5 1 0 0
Screen 2 0 2 3 6 5
Water 11 9 6 1 0 0
Wood (1) 12 7 7 2 0 0
Hood (2) 11 3 10 3 0 0
116
Binary Tests. The parameters are a. b(1..). b(2..).
Texture Name Conservative Rejections at 0.05 level
Acceptances 1 2 _ 3
Brick 5 1 3 7 3 2
Ceiling Tile 12 9 4 2 1 0
Cork 11 6 6 4 0 0
Fur 12 7 8 1 0 0
Grass 13 12 3 1 0 0
Paper 11 8 6 2 0 0
Pebbles 11 , e 6 2 2 0
Sand 14 9 5 2 0 0
Screen 0 0 3 0 2 11
Water 12 7 5 3 1 0
Hood (1) 10 5 8 3 0 0
Wood (2) 10 7 6 1 2 0
117
1292; 1: First-Order (Isotropic). Second-Order
(Anisotropic) Binary Tests. The parameters are a. b(1..).
b(2.1). b(2.2).
Texture Name Conservative Rejections at 0.05 level
Acceptances g 1 2 3 3
Brick 5 3 6 4 3 0
Ceiling Tile 13 10 5 1 0 0
Cork 14 11 5 0 0 0
Fur 10 7 8 1 0 0
Grass 15 11 5 0 0 .0
Paper 13 12 4 0 0 0
Pebbles 11 9 4 0 2 1
Sand 14 12 4 0 0 0
Screen 3 3 1 1 11 0
Water 13 11 4 1 0 0
Wood (1) 13 9 5 2 0 0
Wood (2) _ 12 11 3 1 0 0
118
122;; 8: First-Order (Anisotropic). Second-Order
(Isotropic) Binary Tests. The parameters are a. b(1.1).
b(112)’ b(29o).
Texture Name Conservative Rejections at 0.05 level
Acceptances g 1 2 2 3
Brick 2 0 7 4 3 2
Ceiling Tile 12 11 2 1 2 0
Cork 9 3 6 6 1 0
Fur 9 7 7 2 0 0
Grass 14 9 4 2 1 0
Paper 11 6 4 2 4 0
Pebbles 10 9 2 3 1 1
Sand 12 11 4 1 0 0
Screen 4 2 2 5 1 6
Water 13 7 9 0 0 0
Wood (1) 10 7 6 3 0 0
()1
\D
H
H
D
Wood (2) 13
12212 2 =
119
First-Order
(Anisotropic) Binary Tests.
b(1.2). b(2.1).
Texture Name
Brick
Ceiling Tile
Cork
Fur
Grass
Paper
Pebbles
Sand
Screen
Hater
Hood (1)
Hood (2)
b(292).
Conservative
Acceptances
6
13
14
11
15
12
14
U!
14
14
(Anisotropic).
The parameters
Rejections at
Q .1.
4 7 3 2
10 4 2 o
11 4 1 o
a s 2 o
10 4 2 o
11 4 1' 0
11 2 1 1
14 2 0 0
10 3 3 0
13 2 1 0
Second-Order
are a. b(1.1).
0.05 LEVEL
4
0
0
120
: Best-fitting Binary Texture Models. In the
ow. 'I' means that the fewest rejections were
obtained by using isotropic estimation. while 'A'
signifies the best results were obtained using anisotropic
parameters. The symbol '---' in the second-order column
signifies that the best results were obtained using a
first-order model.
Texture Name L'irst Second Acceptances
70rder Order
Brick I --- , 16
Ceiling Tile I --- 16
Cork I A 14
Fur _ I I 12
Grass A --- 16
Paper I --- 15
Pebbles I --- 14
Sand A A 14
Screen A A 5
Water A A 14
Wood (1) A A 14
Wood (2) A A 14
121
5.5.2 Binomial Texture Results
With the experience gained from the binary fits. we
limited our attention to four samples from each of the
eight gray leVel pictures. First-order models yielded
consistently negative results. The binomial model is
unable to effectively model bimodal or uniform conditional
probabilities. The binomial model always has a peak at
exactly one value for any choice of 6. Also. if there are
two likely gray values which are not contiguous. then no
choice of the e parameter can yield the correct
probabilities.
As in the binary case. third-order analysis cannot be
performed on samples of size 64 by 64 for eight gray level
textures. In the matching experiments. estimation was
performed for third-order textures using 128 by 128
samples. Even this is barely enough. and results in about
thirty percent collapsing of the cells when the cells are
pooled to force expectations of at least one per cell. A
first-order model with anisotropy can have as many as 225
cells. while a full third-order model with anisotropy can
use as many as 225**3 cells. Of course. the number of
cells is limited by the number of points on the lattice.
122
Tables 11 and 12 give the results of fitting
eight-color models to four samples from each texture
class. The format is analagous to Tables 2 through 9.
except that only four samples per texture class were used.
As in the case of the binary textures. good results were
obtained for all but the inhomogeneous textures: Hater.
Hood. and Pebbles. The binomial model has difficulty in
handling large areas of equal brightness. All of the
textures which did not fit the model well are either
blotchy or regular. like the image of the screen.
Fine-grained textures can be handled and. as we shall see.
generated by the binomial model easily. The best results
of the two sets are shown in Table 13.
123
I§b1g 11: Eight gray level. First-order(Anisotropic).
Second-order(Anisotropic). The parameters are a. b(1.1).
b(1.2). b(2.l)9 and b(292).
Texture Name Conservative Rejections at 0.05 level
Acceptances 1 2 1 3
Brick 0 0 0 0 0 4
Ceiling Tile 2 1 3 0 0 0
Cork 4 3 1 0 0 0
Fur 1 1 2 0 1 0
Grass 3 3 1 0 0 0
Paper 4 3 1 0 0 0
Pebbles 0 0 0 0 0 4
Sand 4 4 0 0 0 0
Screen 0 0 0 0 0 4
Water 0 0 0 1 0 1
Hood (1) 1 0 0 0 0 4
Wood (2) 0 0 0 0 0 4
124
12213 12: Eight gray level. First-order(Isotropic).
Second-order(Anisotropic). The parameters are a. b(1..).
b(291). and b(292).
Texture Name Conservative Rejections at 0.05 level
Acceptances 1 2 1 3
Brick 3 0 2 2 0 0
Ceiling Tile 3 3 1 0 0 0
Cork 1 1 0 1 1 1
Fur 4 4 0 0 0 0
Grass 3 1 2 1 0 0
Paper 3 2 2 0 0 0
.Pebbles o o o n o 4
Sand 4 2 2 0 0 0
Screen 0 0 0 0 0 4
Mater 0 0 0 0 0 4
Wood (1) 0 0 0 0 0 4
Hood (2) 0 0 0 0 0 4
125
12913 11: Best-fitting Eight-gray level Texture Models.
In the table below. '1' means that the fewest rejections
were obtained by using isotropic estimation. while 'A'
signifies the best results were obtained using anisotropic
parameterss. The symbol '---' indicates that neither
model was appropriate.
Texture Name First Second Acceptances
Order Order
Brick A I 3
Ceiling Tile A I 3
Cork A A 4
Fur I A 4
Grass A A 4
Paper A A 4
Pebbles --- --- 0
Sand A A 4
Screen --- --- 0
Water --- --- 0
Hood (1) A A 1
Hood (2) --- --- 0
126
5.6 Texture Matching Experiments
This section examines the viability of the Markov
Random Field as a supervised texture generation procedure.
The input texture is measured using the Maximum Likelihood
approach described in section 5.3. The results of that
evaluation are used as the input to the generation
procedure in Figure 12.
As mentioned in section 5.5. the third-order model on
a 64 by 64 texture cannot be reliably fitted since it
causes too many empty or near-empty cells. Limited
experimentation showed that if the image size was
increased to 128 by 128. a sensible chi-square estimate
could be made. though the parameters did not change.very
much from the estimates made on the 64 by 64 textures.
5.6.1 Binary Textures
Binary textures have far simpler structures than
multi-gray level textures. If they are not regular. like
a tiling of the plane by polygons. then we can describe
their general appearance by a few general characteristics.
This is really an abbreviated list of the intuitive
textural attributes defined in the first chapter.
127
(i) Directionality: Vertical. Horizontal. Diagonal.
or None.
(ii) Cluster size: Large. Medium. or Small
(iii) Homogeneity: Homogeneous or Inhomogeneous.
Inhomogeneous images have different characteristics in
different parts of the picture.
(iv) Special Features: Regularity. Line-likeness.
Blob-likeness.
Table 14 illustrates rough characterization on the
basis of one 64 by 64 sample of the features given above.
We can use this guide to evaluate our success in matching
the generated textures. of course. we would like to be
able to say that one texture “looks like" another. but we
need some way of quantifying this correspondence. Figure
29 shows the results of generating the twelve textures
from the estimated parameters. The parameters used to
generate the synthetic textures were obtained by averaging
the parameter estimates from each of the codings of a
single subimage. The choice of subimage was arbitrary. A
third-order estimate was used in all cases except for the
pictures of Mood grain(1) and Pebbles. In these two
128
1221; 13: Binary Texture Characteristics
Cluster
1251252 212211122 Size
Bricks Vertical
Ceiling Vertical
Cork Diagonal
Fur None
Grass Diagonal
Paper Diagonal.
Pebbles None
Sand None
Screen Vertical
Hater Vertical
Nood(1) Horiz.
Nood(2) Horiz.
Large
Small
Small
Large
Small
Small
Large
Medium
Small
Medium
Large
Large
Low
High
High
High
Hihg
Low
High
High
Low
Low
Special
Eeatsces
Regular
None
None
None
None
None
Blobs
None
Lines. dots
None
Blobs
Blobs
129
cases. a first-order model gave better visual results.
First-order models tend to form in blob-like aggregations
and cannot correctly characterize fine-structured
textures. Moreover. they are inadequate in showing any
directionality except vertical and horizontal.
The clear failures are:
Bricks: The regularity and neat rectangles are not
present. Cluster size is close to correct along with the
overall vertical structure. The regularity occurs at an
order of around 20 pixels. and there is little chance of
capturing such a structure from a 64 by 64 picture.
Fur.wood (1). Wood (2): The inhomogeneity is missed in the
synthetic examples. What remains of these pictures after
binary quantization can hardly be called a texture. The
images are estimated to be fine-structured rather than
clean. blob-like shapes like pebbles. because their noise
component modifies the estimate so that third-order
inhibition is used in the generation. This is an
unfortunate consequence of the estimation procedure. It
is possible that a preliminary smoothing might help in
getting a more correct estimate. In all cases. the
directionality is correctly simulated in these textures.
130
The remaining nine textures are reasonably
approximated by the simulated textures. In the pictures
of Cork. Grass. and Paper. diagonality is the overriding
feature and this is correctly modelled. The screen image_
is remarkably similar to the original. The third-order
repulsion effect provides the curious checkerboard effects
along the lines in both the original and the generated
texture. Without this inhibition effect. the image would
resemble the vertical images shown in Figure 16.
131
Eigggg 22 Real and Synthetic Binary Textures: £21 Binary
Bricks 1211 Synthetic Binary Bricks _(_p__)_ Binary Ceiling
tile 5311 Synthetic Binary Ceiling Tile 121 Binary Cork
1511 Synthetic Binary Cork
1g; Binary Fur 5g1; Synthetic
1211 Synthetic
1 Synthetic Binary Paper
Binary
Binary
Grass
519252 22 ...Continwsg! 1.91 Binary Pebbles 19.1.18ynthetic
Binary Pebbles (h) Binary Sand 1511 Synthetic Binary Sand
11; Binary Screen 1111 Synthetic Binary Screen
Figure 22 Continued: 111Binary Water £111Synthetic Binary
Water 121 Binary wood Grain(l) £511 synthetic Binary Hood
Grain(1). 111 Binary Wood Grain(2) 1111 Synthetic Binary
Wood Grain(2).
5.6.2 Binomial Texture Matching
The same basic data set of Brodatz 0173 textures was
used in the binomial matching experiments as was used in
the binary experiments. The pictures were quantized to
have eight gray levels using histogram equalization.
However. during the estimation of the third-order Markov
Random Field parameter set. it was found that far too many
cells were empty or contained only a single member when
the image size was 64 by 64. Therefore. a 128 by 128
image size was used for both estimation and matching. In
all cases. a third-order model was estimated and used to
generate the synthetic textures.
Numerically. the Markov Random Field parameters of
the estimated binomial textures are very similar. which
accounts for the similar appearance of the generated
textures. He have omitted some of the obvious failures:
Water. Hood Grain (1). Hood Grain (2). and Pebbles. Hhen
considered as an eight-gray level image. these textures
take on a distinctly inhomogenous appearance. At this
size. they look like pictures of obiects. whereas the
generated textures look like a fine-grain field. The
Markov Random Field always results in a homogeneous
covering of the image. which cannot result in a blotchy
136
image unless the parameters are extreme. As'an example.
the fur pictures. Figure 30a and 30a1. show the result of
an inhomogeneity in the image.
The other clear failure is the Bricks picture. As in
the binary case. the Bricks image has a regular structure.
The Markov Random Field can only detect a hint of a
vertical structure. This is shown in Figures 30b and
30b1.
Ceiling tile. Cork. Grass. Paper. and Sand are
handled adequately. Missing are the large black holes in
the synthetic ceiling tile picture. but there are some
dense black patches. The distinctly three-dimensional
appearance of the the handmade paper. admittedly a tactile
property. is not captured either.
The Screen texture has the general line-like quality.
but lacks the straightness present in the original. This
could be corrected. but the estimated parameters of the
screen texture do not support the further inhibition
needed to straighten out the lines and keep them narrow
rather than fuzzy.
137
The model seems to be adequate for duplicating the
micro-textures. but is incapable of handling strong
regularity or cloud-like inhomogeneities. These
experiments should be taken as an exploration of the
limits of a purely statistical approach to texture without
any 2 251251 knowledge at all. For example. if we knew
that the Bricks picture was supposed to have rectangles of
a certain size and orientation. then we could start with
them as an outline and then fill in the rectangles with a
Markov Random Field.
138
Real and Synthetic Eight Level Textures 121 Fur
' Fur 121 Bricks 1211 Synthetic Bricks
g Tile 1911 Synthetic Ceiling Tile
519252 39 222112222: 121 Cork 1211 Synthetic Cork
(e) Grass 1e ) Synthetic Grass 111 Paper 1:11 Synthetic
140
Figure 30 ggggigggg: 191 Sand 191) Synthetic Sand
Iii-Ezreeh 1h11 Synthetic Screen
141
5.7 Summary
We have explained the estimation procedure for the
Markov Random Field parameters. Hypothesis testing was
presented as a method of objective evaluation of the fit
between a texture model and a texture. We computed the
fit between Markov Random Field textures and a collection
of samples from the Brodatz texture album [17].
Experiments in the generation of textures that matched
real images were performed. with some successes and
explainable failures.
CHAPTER 6. CO-OCCURRENCE MATRICES
6.1 Introduction
Features based on co-occurrence matrices continue to
play an important role in texture analysis [243. [88]. In
a model-based approach. we would like to determine the
distribution of these features as a function of the model
parameters. If we knew these distributions. then we
could. at least in principle. determine the discriminative
capability of classification algortithms based on the
co-occurrence features. As an example. let 2(1) and 3(2)
be two distinct parameter vectors for a texture model.
Suppose further that the mean. S = EE2:X(i)J. is the same
for textures from either the class defined by 2(1) or the
one defined by 3(2). Then any classification scheme based
solely on S would be unable to discriminate between
textures from the two classes.
A study along these lines was performed Iby Conners
and Harlow [24]. They created textures using a
one-dimensional Markov process. The Markov process
specifies the probability that a pixel with gray level i
follows a pixel with gray level j in a row. The rows of
the image are uncorrelated. This texture model provides a
142
143
method of computing co-occurrence matrices from the
transition matrix of the process. Texture algorithms can
be compared on their capability of distinguishing textures
with different transition matrices.
The difference between the above study and prior
comparative studies [1063 is that Conners's study did not
compare the texture analysis algorithms on the basis of
their performance on a texture database. but rather
predicted their performance on texture classes defined by
model parameters. Empirical studies certainly are of,
interest to users of texture analysis algorithms. but they
guide in the selection of textural recognition algorithms
only insofar as the results of the empirical study can be
generalized to new data.
Our purpose in this chapter is to begin a theoretical
calculation of textural feature distributions based on
Markov Random Field parameters. JWe limit our attention to_
_the case of binary. first-order Markov Random Fields.
144
6.2 Joint Probability Formulation
This section defines the variables we need to
calculate the co-occurrence matrices. We assume that the
binary variables take values in {-1.1}. Following
Bartlett [7]. we have the following expression for the»
joint probability mass function:
(6.1) p(X) = Zexp( -GS(§) -Y(1)U1(§) - Y(2)U2(§))
where
\
(6.2) S(X) = X(r.s)
/
(P95)
\
(6.3) U1(§) = X(r.s)X(r-1.s)
/
(r.s)
\
(6.4) u2(x) : X(r.s)X(r.s-1).
/
145
The lattice is assumed to have a periodic boundary.
The quantity Z in equation (6.1) is called the
partition function. It depends on a. Y(1). and Y(2) and
is a normalizing constant .30 that p(X) is a valid
probability mass function.
The quantity EES(§)] is called the mean of the
process. If the value of a is zero. then EES(§)] is also
zero. This is not hard to show. For each coloring 5. let
-1 denote the negative of 5. in the sense that the value
of each point X(r.s) of the lattice L is reversed in sign.
Then my = -s<-x>. but 01(1) = U1(-§). and 02(5) =
U2(-5). Since the quantity a is zero. we see from
equation (6.1) that p(X) = p(-X). Hence. we conclude that
Ets