Search results

Title: A framework for combining ancillary information with primary biometric traits
Creator: Ding, Yaohui
Date: 2018
Collection: Electronic Theses & Dissertations
Description: "Biometric systems recognize individuals based on their biological attributes such as faces, fingerprints and iris. However, in several scenarios, additional ancillary information such as the biographic and demographic information of a user (e.g., name, gender, age, ethnicity), or the image quality of the biometric sample, anti-spoofing measurements, etc. may be available. While previous literature has studied the impact of such ancillary information on biometric system performance, there is...
Show more"Biometric systems recognize individuals based on their biological attributes such as faces, fingerprints and iris. However, in several scenarios, additional ancillary information such as the biographic and demographic information of a user (e.g., name, gender, age, ethnicity), or the image quality of the biometric sample, anti-spoofing measurements, etc. may be available. While previous literature has studied the impact of such ancillary information on biometric system performance, there is limited work on systematically incorporating them into the biometric matching framework. In this dissertation, we develop a principled framework to combine ancillary information with biometric match scores. The incorporation of ancillary information raises several challenges. Firstly, ancillary information such as gender, ethnicity and other demographic attributes lack distinctiveness and can be used to distinguish population groups rather than individuals. Secondly, ancillary information such as image quality and anti-spoof measurements may have different numerical ranges and interpretations. Further, most of the ancillary information cannot be automatically extracted without errors. Even the direct collection of ancillary information from subjects may be susceptible to transcription errors (e.g., errors in entering the data). Thirdly, the relationships between ancillary attributes and biometric traits may not be evident. In this regard, this dissertation makes three contributions. The first contribution entails the design of a Bayesian Belief Network (BBN) to model the relationship between biometric scores and ancillary factors, and exploiting the ensuing structure in a fusion framework. The ancillary information considered by the network includes image quality and anti-spoof measures. Experiments convey the importance of explicitly incorporating such information in a biometric system. The second contribution is the design of a Generalized Additive Model (GAM) that uses spline functions to model the correlation between match scores and ancillary attributes, and then learns a transformation function to normalize the match scores prior to fusion. The resulting framework can also be used to predict in advance if fusing match scores with certain demographic attributes is beneficial in the context of a specific biometric matcher. Experiments indicate that the proposed method can be used to significantly improve the recognition accuracy of state-of-the-art face matchers. The third contribution is the design of an ensemble of One Class Support Vector Machines (OC-SVMs) to combine multiple anti-spoofing measurements in order to mitigate the concerns associated with the issue of "imbalanced training sets" and "insufficient spoof samples" encountered by conventional anti-spoofing algorithms. In the proposed method, the spoof detection problem is formulated as a one-class problem, where the focus is on modeling a real fingerprint using multiple feature sets. The one-class classifiers corresponding to these multiple feature sets are then combined to generate a single classifier for spoof detection. Experimental results convey the importance of this technique in detecting spoofs made of materials that were not included in the training data. In summary, this dissertation seeks to advance our understanding of systematically exploiting ancillary information in designing effective biometric recognition systems by developing and evaluating multiple statistical models."--Pages ii-iii.
Show less

Title: Low rank models for multi-dimensional data recovery and image super-resolution
Creator: Al-Qizwini, Mohammed
Date: 2017
Collection: Electronic Theses & Dissertations
Description: "In the past decade tremendous research efforts focused on signals with specific features, especially sparse and low rank signals. Researchers showed that these signals can be recovered from much smaller number of samples than the Nyquist rate. These efforts were promising for several applications in which the nature of the data is known to be sparse or low rank, but the available samples are much fewer than what is required by the traditional signal processing algorithms to grant an exact...
Show more"In the past decade tremendous research efforts focused on signals with specific features, especially sparse and low rank signals. Researchers showed that these signals can be recovered from much smaller number of samples than the Nyquist rate. These efforts were promising for several applications in which the nature of the data is known to be sparse or low rank, but the available samples are much fewer than what is required by the traditional signal processing algorithms to grant an exact recovery. Our objective in the first part of this thesis is to develop new algorithms for low rank data recovery from few observed samples and for robust low rank and sparse data separation using the Robust Principal Component Analysis (RPCA). Most current approaches in this class of algorithms are based on using the computationally expensive Singular Value Decomposition (SVD) in each iteration to minimize the nuclear norm. In particular, we first develop new algorithms for low rank matrix completion that are more robust to noise and converge faster than the previous algorithms. Furthermore, we generalize our recovery function to the multi-dimensional tensor domain to target the applications that deal with multi-dimensional data. Based on this generalized function, we propose a new tensor completion algorithm to recover multi-dimensional tensors from few observed samples. We also used the same generalized functions for robust tensor recovery to reconstruct the sparse and low rank tensors from the tensor that is formed by the superposition of those parts. The experimental results for this application showed that our algorithms provide comparable performance, or even outperforms, state-of-the-art matrix completion, tensor completion and robust tensor recovery algorithms; but at the same time our algorithms converge faster. The main objective of the second part of the thesis develops new algorithms for example based single image super-resolution. In this type of applications, we observe a low-resolution image and using some external "example" high-resolution - low-resolution images pairs, we recover the underlying high-resolution image. The previous efforts in this field either assumed that there is a one-to-one mapping between low-resolution and high-resolution image patches or they assumed that the high-resolution patches span the lower dimensional space. In this thesis, we propose a new algorithm that parts away from these assumptions. Our algorithm uses a subspace similarity measure to find the closes high-resolution patch to each low-resolution patch. The experimental results showed that DMCSS achieves clear visual improvements and an average of 1dB improvement in PSNR over state-of-the-art algorithms in this field. Under this thesis, we are currently pursuing other low rank and image super-resolution applications to improve the performance of our current algorithms and to find other algorithms that can run faster and perform even better."--Pages ii-iii.
Show less

Title: Large-scale high dimensional distance metric learning and its application to computer vision
Creator: Qian, Qi
Date: 2015
Collection: Electronic Theses & Dissertations
Description: Learning an appropriate distance function (i.e., similarity) is one of the key tasks in machine learning, especially for distance based machine learning algorithms, e.g., $k$-nearest neighbor classifier, $k$-means clustering, etc. Distance metric learning (DML), the subject to be studied in this dissertation, is designed to learn a metric that pulls the examples from the same class together and pushes the examples from different classes away from each other. Although many DML algorithms have...
Show moreLearning an appropriate distance function (i.e., similarity) is one of the key tasks in machine learning, especially for distance based machine learning algorithms, e.g., $k$-nearest neighbor classifier, $k$-means clustering, etc. Distance metric learning (DML), the subject to be studied in this dissertation, is designed to learn a metric that pulls the examples from the same class together and pushes the examples from different classes away from each other. Although many DML algorithms have been developed in the past decade, most of them can handle only small data sets with hundreds of features, significantly limiting their applications to real world applications that often involve millions of training examples represented by hundreds of thousands of features. Three main challenges are encountered to learn the metric from these large-scale high dimensional data: (i) To make sure that the learned metric is a Positive Semi-Definitive (PSD) matrix, a projection into the PSD cone is required at every iteration, whose cost is cubic in the dimensionality making it unsuitable for high dimensional data; (ii) The number of variables that needs to be optimized in DML is quadratic in the dimensionality, which results in the slow convergence rate in optimization and high requirement of memory storage; (iii) The number of constraints used by DML is at least quadratic, if not cubic, in the number of examples depending on if pairwise constraints or triplet constraints are used in DML. Besides, features can be redundant due to high dimensional representations (e.g., face features) and DML with feature selection is preferred for these applications.The main contribution of this dissertation is to address these challenges both theoretically and empirically. First, for the challenge arising from the PSD projection, we exploit the mini-batch strategy and adaptive sampling with smooth loss function to significantly reduce the number of updates (i.e., projections) while keeping the similar performance. Second, for the challenge arising from high dimensionality, we propose a dual random projection approach, which enjoys the light computation due to the usage of random projection and at the same time, significantly improves the effectiveness of random projection. Third, for the challenge with large-scale constraints, we develop a novel multi-stage metric learning framework. It divides the original optimization problem into multiple stages. It reduces the computation by adaptively sampling a small subset of constraints at each stage. Finally, to handle redundant features with group property, we develop a greedy algorithm that selects feature group and learns the corresponding metric simultaneously at each iteration leading to further improvement of learning efficiency when combined with adaptive mini-batch strategy and incremental sampling. Besides the theoretical and empirical investigation of DML on the benchmark datasets of machine learning, we also apply the proposed methods to several important computer vision applications (i.e., fine-grained visual categorization (FGVC) and face recognition).
Show less

Title: Novel learning algorithms for mining geospatial data
Creator: Yuan, Shuai (Software engineer)
Date: 2017
Collection: Electronic Theses & Dissertations
Description: Geospatial data have a wide range of applicability in many disciplines, including environmental science, urban planning, healthcare, and public administration. The proliferation of such data in recent years have presented opportunities to develop novel data mining algorithms for modeling and extracting useful patterns from the data. However, there are many practical issues remain that must be addressed before the algorithms can be successfully applied to real-world problems. First, the...
Show moreGeospatial data have a wide range of applicability in many disciplines, including environmental science, urban planning, healthcare, and public administration. The proliferation of such data in recent years have presented opportunities to develop novel data mining algorithms for modeling and extracting useful patterns from the data. However, there are many practical issues remain that must be addressed before the algorithms can be successfully applied to real-world problems. First, the algorithms must be able to incorporate spatial relationships and other domain constraints defined by the problem. Second, the algorithms must be able to handle missing values, which are common in many geospatial data sets. In particular, the models constructed by the algorithms may need to be extrapolated to locations with no observation data. Another challenge is to adequately capture the nonlinear relationship between the predictor and response variables of the geospatial data. Accurate modeling of such relationship is not only a challenge, it is also computationally expensive. Finally, the variables may interact at different spatial scales, making it necessary to develop models that can handle multi-scale relationships present in the geospatial data. This thesis presents the novel algorithms I have developed to overcome the practical challenges of applying data mining to geospatial datasets. Specifically, the algorithms will be applied to both supervised and unsupervised learning problems such as cluster analysis and spatial prediction. While the algorithms are mostly evaluated on datasets from the ecology domain, they are generally applicable to other geospatial datasets with similar characteristics. First, a spatially constrained spectral clustering algorithm is developed for geospatial data. The algorithm provides a flexible way to incorporate spatial constraints into the spectral clustering formulation in order to create regions that are spatially contiguous and homogeneous. It can also be extended to a hierarchical clustering setting, enabling the creation of fine-scale regions that are nested wholly within broader-scale regions. Experimental results suggest that the nested regions created using the proposed approach are more balanced in terms of their sizes compared to the regions found using traditional hierarchical clustering methods. Second, a supervised hash-based feature learning algorithm is proposed for modeling nonlinear relationships in incomplete geospatial data. The proposed algorithm can simultaneously infer missing values while learning a small set of discriminative, nonlinear features of the geospatial data. The efficacy of the algorithm is demonstrated using synthetic and real-world datasets. Empirical results show that the algorithm is more effective than the standard approach of imputing the missing values before applying nonlinear feature learning in more than 75% of the datasets evaluated in the study. Third, a multi-task learning framework is developed for modeling multiple response variables in geospatial data. Instead of training the local models independently for each response variable at each location, the framework simultaneously fits the local models for all response variables by optimizing a joint objective function with trace-norm regularization. The framework also leverages the spatial autocorrelation between locations as well as the inherent correlation between response variables to improve prediction accuracy. Finally, a multi-level, multi-task learning framework is proposed to effectively train predictive models from nested geospatial data containing predictor variables measured at multiple spatial scales. The framework enables distinct models to be developed for each coarse- scale region using both its fine-level and coarse-level features. It also allows information to be shared among the models through a common set of latent features. Empirical results show that such information sharing helps to create more robust models especially for regions with limited or no training data. Another advantage of using the multi-level, multi-task learning framework is that it can automatically identify potential cross-scale interactions between the regional and local variables.
Show less

Title: Automated Speaker Recognition in Non-ideal Audio Signals Using Deep Neural Networks
Creator: Chowdhury, Anurag
Date: 2021
Collection: Electronic Theses & Dissertations
Description: Speaker recognition entails the use of the human voice as a biometric modality for recognizing individuals. While speaker recognition systems are gaining popularity in consumer applications, most of these systems are negatively affected by non-ideal audio conditions, such as audio degradations, multi-lingual speech, and varying duration audio. This thesis focuses on developing speaker recognition systems robust to non-ideal audio conditions.Firstly, a 1-Dimensional Convolutional Neural...
Show moreSpeaker recognition entails the use of the human voice as a biometric modality for recognizing individuals. While speaker recognition systems are gaining popularity in consumer applications, most of these systems are negatively affected by non-ideal audio conditions, such as audio degradations, multi-lingual speech, and varying duration audio. This thesis focuses on developing speaker recognition systems robust to non-ideal audio conditions.Firstly, a 1-Dimensional Convolutional Neural Network (1D-CNN) is developed to extract noise-robust speaker-dependent speech characteristics from the Mel Frequency Cepstral Coefficients (MFCC). Secondly, the 1D-CNN-based approach is extended to develop a triplet-learning-based feature-fusion framework, called 1D-Triplet-CNN, for improving speaker recognition performance by judiciously combining MFCC and Linear Predictive Coding (LPC) features. Our hypothesis rests on the observation that MFCC and LPC capture two distinct aspects of speech: speech perception and speech production. Thirdly, a time-domain filterbank called DeepVOX is learned from vast amounts of raw speech audio to replace commonly-used hand-crafted filterbanks, such as the Mel filterbank, in speech feature extractors. Finally, a vocal style encoding network called DeepTalk is developed to learn speaker-dependent behavioral voice characteristics to improve speaker recognition performance. The primary contribution of the thesis is the development of deep learning-based techniques to extract discriminative, noise-robust physical and behavioral voice characteristics from non-ideal speech audio. A large number of experiments conducted on the TIMIT, NTIMIT, SITW, NIST SRE (2008, 2010, and 2018), Fisher, VOXCeleb, and JukeBox datasets convey the efficacy of the proposed techniques and their importance in improving speaker recognition performance in non-ideal audio conditions.
Show less

Title: Iris Recognition : Enhancing Security and Improving Performance
Creator: Sharma, Renu
Date: 2022
Collection: Electronic Theses & Dissertations
Description: Biometric systems recognize individuals based on their physical or behavioral traits, viz., face, iris, and voice. Iris (the colored annular region around the pupil) is one of the most popular biometric traits due to its uniqueness, accuracy, and stability. However, its widespread usage raises security concerns against various adversarial attacks. Another challenge is to match iris images with other compatible biometric modalities (i.e., face) to increase the scope of human identification....
Show moreBiometric systems recognize individuals based on their physical or behavioral traits, viz., face, iris, and voice. Iris (the colored annular region around the pupil) is one of the most popular biometric traits due to its uniqueness, accuracy, and stability. However, its widespread usage raises security concerns against various adversarial attacks. Another challenge is to match iris images with other compatible biometric modalities (i.e., face) to increase the scope of human identification. Therefore, the focus of this thesis is two-fold: firstly, enhance the security of the iris recognition system by detecting adversarial attacks, and secondly, accentuate its performance in iris-face matching.To enhance the security of the iris biometric system, we work over two types of adversarial attacks - presentation and morph attacks. A presentation attack (PA) occurs when an adversary presents a fake or altered biometric sample (plastic eye, cosmetic contact lens, etc.) to a biometric system to obfuscate their own identity or impersonate another identity. We propose three deep learning-based iris PA detection frameworks corresponding to three different imaging modalities, namely NIR spectrum, visible spectrum, and Optical Coherence Tomography (OCT) imaging inputting a NIR image, visible-spectrum video, and cross-sectional OCT image, respectively. The techniques perform effectively to detect known iris PAs as well as generalize well across unseen attacks, unseen sensors, and multiple datasets. We also presented the explainability and interpretability of the results from the techniques. Our other focuses are robustness analysis and continuous update (retraining) of the trained iris PA detection models. Another burgeoning security threat to biometric systems is morph attacks. A morph attack entails the generation of an image (morphed image) that embodies multiple different identities. Typically, a biometric image is associated with a single identity. In this work, we first demonstrate the vulnerability of iris recognition techniques to morph attacks and then develop techniques to detect the morphed iris images.The second focus of the thesis is to improve the performance of a cross-modal system where iris images are matched against face images. Cross-modality matching involves various challenges, such as cross-spectral, cross-resolution, cross-pose, and cross-temporal. To address these challenges, we extract common features present in both images using a multi-channel convolutional network and also generate synthetic data to augment insufficient training data using a dual-variational autoencoder framework. The two focus areas of this thesis improve the acceptance and widespread usage of the iris biometric system.
Show less

Title: Discrete de Rham-Hodge Theory
Creator: Zhao, Rundong
Date: 2020
Collection: Electronic Theses & Dissertations
Description: We present a systematic treatment to 3D shape analysis based on the well-established de Rham-Hodge theory in differential geometry and topology. The computational tools we developed are widely applicable to research areas such as computer graphics, computer vision, and computational biology. We extensively tested it in the context of 3D structure analysis of biological macromolecules to demonstrate the efficacy and efficiency of our method in potential applications. Our contributions are...
Show moreWe present a systematic treatment to 3D shape analysis based on the well-established de Rham-Hodge theory in differential geometry and topology. The computational tools we developed are widely applicable to research areas such as computer graphics, computer vision, and computational biology. We extensively tested it in the context of 3D structure analysis of biological macromolecules to demonstrate the efficacy and efficiency of our method in potential applications. Our contributions are summarized in the following aspects. First, we present a compendium of discrete Hodge decompositions of vector fields, which provides the primary building block of the de Rham-Hodge theory for computations performed on the commonly used tetrahedral meshes embedded in the 3D Euclidean space. Second, we present a real-world application of the above computational tool to 3D shape analysis on biological macromolecules. Finally, we extend the above method to an evolutionary de Rham-Hodge method to provide a unified paradigm for the multiscale geometric and topological analysis of evolving manifolds constructed from a filtration, which induces a family of evolutionary de Rham complexes. Our work on the decomposition of vector fields, spectral shape analysis on static shapes, and evolving shapes has already shown its effectiveness in biomolecular applications and will lead to a rich set of features for machine learning-based shape analysis currently under development.
Show less

Title: SIGN LANGUAGE RECOGNIZER FRAMEWORK BASED ON DEEP LEARNING ALGORITHMS
Creator: Akandeh, Atra
Date: 2021
Collection: Electronic Theses & Dissertations
Description: According to the World Health Organization (WHO, 2017), 5% of the world’s population have hearing loss. Most people with hearing disabilities communicate via sign language, which hearing people find extremely difficult to understand. To facilitate communication of deaf and hard of hearing people, developing an efficient communication system is a necessity. There are many challenges associated with the Sign Language Recognition (SLR) task, namely, lighting conditions, complex background,...
Show moreAccording to the World Health Organization (WHO, 2017), 5% of the world’s population have hearing loss. Most people with hearing disabilities communicate via sign language, which hearing people find extremely difficult to understand. To facilitate communication of deaf and hard of hearing people, developing an efficient communication system is a necessity. There are many challenges associated with the Sign Language Recognition (SLR) task, namely, lighting conditions, complex background, signee body postures, camera position, occlusion, complexity and large variations in hand posture, no word alignment, coarticulation, etc.Sign Language Recognition has been an active domain of research since the early 90s. However, due to computational resources and sensing technology constraints, limited advancement has been achieved over the years. Existing sign language translation systems mostly can translate a single sign at a time, which makes them less effective in daily-life interaction. This work develops a novel sign language recognition framework using deep neural networks, which directly maps videos of sign language sentences to sequences of gloss labels by emphasizing critical characteristics of the signs and injecting domain-specific expert knowledge into the system. The proposed model also allows for combining data from variant sources and hence combating limited data resources in the SLR field.
Show less

Title: Human activity monitoring by smart devices
Creator: Bi, Chongguang
Date: 2020
Collection: Electronic Theses & Dissertations
Description: The topic of the Internet-of-Things (IoT) has been discussed and studied extensively since 2010. It provides various solutions for enhancing the user's experience, monitoring the user's behaviors, and improving the lifestyle. With careful design, these systems can be built with off-the-shelf smartphones and wearables. The detected result can be used as feedback for the user to understand his/her behavior, improve the lifestyle, or avoid the danger. Furthermore, the result also provides a...
Show moreThe topic of the Internet-of-Things (IoT) has been discussed and studied extensively since 2010. It provides various solutions for enhancing the user's experience, monitoring the user's behaviors, and improving the lifestyle. With careful design, these systems can be built with off-the-shelf smartphones and wearables. The detected result can be used as feedback for the user to understand his/her behavior, improve the lifestyle, or avoid the danger. Furthermore, the result also provides a valuable data source for the studies in psychology and sociology.However, designing an IoT system to monitor human activities is challenging due to multiple factors. Some systems require high computing capability or a long time of data collection; some systems must detect some specific behaviors as quickly as possible in real-time; some systems suffer constant and irregular noise. In order to address these challenges, the designer must carefully consider the use case of the IoT system and select proper machine learning algorithms. This dissertation shows three designs of the IoT systems for the improvement of family mealtime experience and driving safety. The procedure for each design is introduced in detail, including the architecture of the system, the selection of features, and the evaluation of algorithms. From the case studies in this dissertation, several special aspects of monitoring human activities are discovered. Since human activity is strongly related to the time-series and may change along time, the algorithm should be sensitive to context, be adaptive to dynamic conditions, process readable features, and benefit directly from prior knowledge. This discovery will serve as a guide about how to analyze and solve a problem with the IoT systems in the future.
Show less

Title: A robust method for addressing pupil dilation in iris recognition
Creator: Pasula, Raghunandan
Date: 2016
Collection: Electronic Theses & Dissertations
Description: "The rich texture of the iris is being used as a biometric cue in several human recognition systems. Iris recognition systems are fairly robust to small changes in illumination and pose. However there are a number of factors that still adversely affect the performance of an iris matcher. These include occlusion, large deviation in gaze, low image resolution, long acquisition distance and pupil dilation. Large differences in pupil size increases the dissimilarity between iris images of the...
Show more"The rich texture of the iris is being used as a biometric cue in several human recognition systems. Iris recognition systems are fairly robust to small changes in illumination and pose. However there are a number of factors that still adversely affect the performance of an iris matcher. These include occlusion, large deviation in gaze, low image resolution, long acquisition distance and pupil dilation. Large differences in pupil size increases the dissimilarity between iris images of the same eye. In this work, the degradation of match scores due to pupil dilation is systematically studied using Hamming Distance histograms. A novel rule-based fusion technique based on the aforementioned study is proposed to alleviate the effect of pupil dilation. The proposed method computes a new distance score at every pixel location based on the similarities between IrisCode bits that were generated using Gabor Filters at different resolutions. Experiments show that the proposed method increases the genuine accept rate from 76% to 90% at 0.0001% false accept rate when comparing images with large differences in pupil sizes in the WVU-PLR dataset. The proposed method is also shown to improve the performance of iris recognition on other non-ideal iris datasets. In summary, the use of multi-resolution Gabor Filters in conjunction with a rule-based integration of decisions at the pixel (bit) level is observed to improve the resilience of iris recognition to differences in pupil size."--Page ii.
Show less

Title: COLLABORATIVE DISTRIBUTED DEEP LEARNING SYSTEMS ON THE EDGES
Creator: Zeng, Xiao
Date: 2021
Collection: Electronic Theses & Dissertations
Description: Deep learning has revolutionized a wide range of fields. In spite of its success, most deep learning systems are proposed in the cloud, where data are processed in a centralized manner with abundant compute and network resources. This raises a problem when deep learning is deployed on the edge where distributed compute resources are limited. In this dissertation, we propose three distributed systems to enable collaborative deep learning on the edge. These three systems target different...
Show moreDeep learning has revolutionized a wide range of fields. In spite of its success, most deep learning systems are proposed in the cloud, where data are processed in a centralized manner with abundant compute and network resources. This raises a problem when deep learning is deployed on the edge where distributed compute resources are limited. In this dissertation, we propose three distributed systems to enable collaborative deep learning on the edge. These three systems target different scenarios and tasks. The first system dubbed Distream is a distributed live video analytics system based on the smart camera-edge cluster architecture. Distream fully utilizes the compute resources at both ends to achieve optimized system performance. The second system dubbed Mercury is a system that addresses the key bottleneck of collaborative learning. Mercury enhances the training efficiency of on-device collaborative learning without compromising the accuracies of the trained models. The third system dubbed FedAce is a distributed training system that improves training efficiency under federated learning setting where private on-device data are not allowed to be shared among local devices. Within each participating client, FedAce achieves such improvement by prioritizing important data. In the server where model aggregation is performed, FedAce exploits the client importance and prioritizes important clients to reduce stragglers and reduce the total number of rounds. In addition, FedAce conducts federated model compression to reduce the per-round communication cost and obtains a compact model after training completes. Extensive experiments show that the proposed three systems are able to achieve significant improvements over status-quo systems.
Show less

Title: Novel Depth Representations for Depth Completion with Application in 3D Object Detection
Creator: Imran, Saif Muhammad
Date: 2022
Collection: Electronic Theses & Dissertations
Description: Depth completion refers to interpolating a dense, regular depth grid from sparse and irregularly sampled depth values, often guided by high-resolution color imagery. The primary goal of depth completion is to estimate depth. In practice methods are trained by minimizing an error between predicted dense depth and ground-truth depth, and are evaluated by how well they minimize this error. Here we identify a second goal which is to avoid smearing depth across depth discontinuities. This second...
Show moreDepth completion refers to interpolating a dense, regular depth grid from sparse and irregularly sampled depth values, often guided by high-resolution color imagery. The primary goal of depth completion is to estimate depth. In practice methods are trained by minimizing an error between predicted dense depth and ground-truth depth, and are evaluated by how well they minimize this error. Here we identify a second goal which is to avoid smearing depth across depth discontinuities. This second goal is important because it can improve downstream applications of depth completion such as object detection and pose estimation. However, we also show that the goal of minimizing error can conflict with the goal of eliminating depth smearing.In this thesis, we propose two novel representations of depths that can encode depth discontinuity across object surfaces by allowing multiple depth estimation in the spatial domain. In order to learn these new representations, we propose carefully designed loss functions and show their effectiveness in deep neural network learning. We show how our representations can avoid inter-object depth mixing and also beat state of the art metrics for depth completion. The quality of ground-truth depth in real-world depth completion problems is another key challenge for learning and accurate evaluation of methods. Ground truth depth created from semi-automatic methods suffers from sparse sampling and errors at object boundaries. We show that the combination of these errors and the commonly used evaluation measure has promoted solutions that mix depths across boundaries in current methods. The thesis proposes alternate depth completion performance measures that reduce preference for mixed depths and promote sharp boundaries.The thesis also investigates whether additional points from depth completion methods can help in a challenging and high-level perception problem; 3D object detection. It shows the effect of different depth noises originated from depth estimates on detection performances and proposes some effective ways to reduce noise in the estimate and overcome architecture limitations. The method is demonstrated on both real-world and synthetic datasets.
Show less

Title: Adaptive on-device deep learning systems
Creator: Fang, Biyi
Date: 2019
Collection: Electronic Theses & Dissertations
Description: "Mobile systems such as smartphones, drones, and augmented-reality headsets are revolutionizing our lives. On-device deep learning is regarded as the key enabling technology for realizing their full potential. This is because communication with cloud adds additional latency or cost, or the applications must operate even with intermittent internet connectivity.The key to achieving the full promise of these mobile vision systems is effectively analyzing the streaming video frames. However,...
Show more"Mobile systems such as smartphones, drones, and augmented-reality headsets are revolutionizing our lives. On-device deep learning is regarded as the key enabling technology for realizing their full potential. This is because communication with cloud adds additional latency or cost, or the applications must operate even with intermittent internet connectivity.The key to achieving the full promise of these mobile vision systems is effectively analyzing the streaming video frames. However, processing streaming video frames taken in mobile settings is challenging in two folds. First, the processing usually involves multiple computer vision tasks. This multi-tenant characteristic requires mobile vision systems to concurrently run multiple applications that target different vision tasks. Second, the context in mobile settings can be frequently changed. This requires mobile vision systems to be able to switch applications to execute new vision tasks encountered in the new context.In this article, we fill this critical gap by proposing NestDNN, a framework that enables resource-aware multi-tenant on-device deep learning for continuous mobile vision. NestDNN enables each deep learning model to offer flexible resource-accuracy trade-offs. At runtime,it dynamically selects the optimal resource-accuracy trade-off for each deep learning model to fit the model's resource demand to the system's available runtime resources. In doing so, NestDNN efficiently utilizes the limited resources in mobile vision systems to jointly maximize the performance of all the concurrently running applications.Although NestDNN is able to efficiently utilize the resource by being resource-aware, it essentially treats the content of each input image equally and hence does not realize the full potential of such pipelines. To realize its full potential, we further propose FlexDNN, a novel content-adaptive framework that enables computation-efficient DNN-based on-device video stream analytics based on early exit mechanism. Compared to state-of-the-art earlyexit-based solutions, FlexDNN addresses their key limitations and pushes the state-of-the-artforward through its innovative fine-grained design and automatic approach for generating the optimal network architecture."--Pages ii-iii.
Show less

Title: An accurate, efficient, and robust fingerprint presentation attack detector
Creator: Chugh, Tarang
Date: 2020
Collection: Electronic Theses & Dissertations
Description: The individuality and persistence of fingerprints is being leveraged for a plethora of day-to-day automated person recognition applications, ranging from social benefits disbursements and unlocking smartphones to law enforcement and border security. While the primary purpose of a fingerprint recognition system is to ensure reliable and accurate user recognition, the security of the system itself can be jeopardized by the use of fingerprint presentation attacks (PAs). A fingerprint PA is...
Show moreThe individuality and persistence of fingerprints is being leveraged for a plethora of day-to-day automated person recognition applications, ranging from social benefits disbursements and unlocking smartphones to law enforcement and border security. While the primary purpose of a fingerprint recognition system is to ensure reliable and accurate user recognition, the security of the system itself can be jeopardized by the use of fingerprint presentation attacks (PAs). A fingerprint PA is defined as a presentation, of a spoof (fake), altered, or cadaver finger, to the data capture system (fingerprint reader) intended to interfere with the recording of the true fingerprint sample/identity, and thereby preventing correct user recognition.In this thesis, we present an automated, accurate, and reliable software-only fingerprint presentation attack detector (PAD), called Fingerprint Spoof Buster. Specifically, we propose a deep convolutional neural network (CNN) based approach utilizing local patches centered and aligned using fingerprint minutiae. The proposed PAD achieves state-of-the-art performance on publicly available liveness detection databases (LivDet) and large-scale government controlled tests as part of the IARPA ODIN program. Additionally, we present a graphical user interface that highlights local regions of the fingerprint image as bonafide or PA for visual examination. This offers significant advantage over existing PAD solutions that rely on a single spoof score for the entire fingerprint image.Deep learning-based solutions are infamously resource intensive (both memory and processing) and require special hardware such as graphical processing units (GPUs). With the goal of real-time inference in low-resource environments, such as smartphones and embedded devices, we propose a series of optimizations including simplifying the network architecture and quantizing model weights (for byte computations instead of floating point arithmetic). These optimizations enabled us to develop a light-weight version of the PAD, called Fingerprint Spoof Buster Lite, as an Android application, which can execute on a commodity smartphone (Samsung Galaxy S8) with a minimal drop in PAD performance (from TDR = 95.7% to 95.3% FDR = 0.2%) in under 100ms.Typically, deep learning-based solutions are considered as "black-box" systems due to the lack of interpretability of their decisions. One of the major limitations of the existing PAD solutions is their poor generalization against PA materials not seen during training. While it is observed that some materials are easier to detect (e.g. EcoFlex) compared to others (e.g. Silgum) when left out from training, the underlying reasons are unknown. We present a framework to understand and interpret the generalization (cross-material) performance of the proposed PAD by investigating the material properties and visualizing the bonafide and PA samples in the multidimensional feature space learned by deep networks. Furthermore, we present two different approaches to improve the generalization performance: (i) a style transfer-based wrapper, called Universal Material Generator (UMG), and (ii) a dynamic approach utilizing temporal analysis of a sequence of fingerprint image frames. The two proposed approaches are shown to significantly improve the generalization performance evaluated on large databases of bonafide and PA samples.Lastly, fingerprint readers based on conventional imaging technologies, such as optical, capacitive, and thermal, only image the 2D surface fingerprint making them an easy target for presentation attacks. In contrast, Optical Coherent Tomography (OCT) imaging technology provides rich depth information, including the internal fingerprint, eccrine (sweat) glands, as well as PA instruments (spoofs) placed over finger skin. As a final contribution, we present an automated PAD approach utilizing cross-sectional OCT depth profile scans which is shown to achieve a TDR of 99.73% FDR of 0.2% on a database of 3,413 bonafide and 357 PA OCT scans, fabricated using 8 different PA materials. We also identify the crucial regions in the OCT scans necessary for PA detection.
Show less

Title: Orientation guided texture synthesis using PatchMatch
Creator: Dutka, Rosemary L.
Date: 2013
Collection: Electronic Theses & Dissertations
Description: Texture describes the unique structural patterns that we perceive in the world. Various surface geometric details such as animal fur, plant leaves, and carpets can be thought of as texture. In computer graphics, textures stored as images are ubiquitously used to decorate boundary surfaces of objects. There are multiple approaches to acquire realistic and aesthetically pleasing textures. One of the most popular methods is a process known as texture synthesis, in which we produce seamless...
Show moreTexture describes the unique structural patterns that we perceive in the world. Various surface geometric details such as animal fur, plant leaves, and carpets can be thought of as texture. In computer graphics, textures stored as images are ubiquitously used to decorate boundary surfaces of objects. There are multiple approaches to acquire realistic and aesthetically pleasing textures. One of the most popular methods is a process known as texture synthesis, in which we produce seamless nonrepetitive textures from a small patch of texture sample.In this thesis, we present an orientation guided fast texture synthesis based on an image editing tool, PatchMatch, which is included in PhotoShop. Given an example image, our model adopts a hierarchical process to improve retention of structural texture features at multiple scales. We generalize PatchMatch by using orientation to guide the alignment of texture features, indicated by a planar direction field, in the creation of large texture patches. To demonstrate the effectiveness of our approach, we first apply our algorithm in designing new textures with two and four-way symmetry which can be extended to n-way symmetry, and then in enhancing latent fingerprints. Furthermore, our results show empirically that orientation guided PatchMatch has the advantages of providing control over the density of singularities without knowing the exact locations and reducing spurious singularities.
Show less

Title: Towards a Robust Unconstrained Face Recognition Pipeline with Deep Neural Networks
Creator: Shi, Yichun
Date: 2021
Collection: Electronic Theses & Dissertations
Description: Face recognition is a classic problem in the field of computer vision and pattern recognition due to its wide applications in real-world problems such as access control, identity verification, physical security, surveillance, etc. Recent progress in deep learning techniques and the access to large-scale face databases has lead to a significant improvement of face recognition accuracy under constrained and semi-constrained scenarios. Deep neural networks are shown to surpass human performance...
Show moreFace recognition is a classic problem in the field of computer vision and pattern recognition due to its wide applications in real-world problems such as access control, identity verification, physical security, surveillance, etc. Recent progress in deep learning techniques and the access to large-scale face databases has lead to a significant improvement of face recognition accuracy under constrained and semi-constrained scenarios. Deep neural networks are shown to surpass human performance on Labeled Face in the Wild (LFW), which consists of celebrity photos captured in the wild. However, in many applications, e.g. surveillance videos, where we cannot assume that the presented face is under controlled variations, the performance of current DNN-based methods drop significantly. The main challenges in such an unconstrained face recognition problem include, but are not limited to: lack of labeled data, robust face normalization, discriminative representation learning and the ambiguity of facial features caused by information loss.In this thesis, we propose a set of methods that attempt to address the above challenges in unconstrained face recognition systems. Starting from a classic deep face recognition pipeline, we review how each step in this pipeline could fail on low-quality uncontrolled input faces, what kind of solutions have been studied before, and then introduce our proposed methods. The various methods proposed in this thesis are independent but compatible with each other. Experiment on several challenging benchmarks, e.g. IJB-C and IJB-S show that the proposed methods are able to improve the robustness and reliability of deep unconstrained face recognition systems. Our solution achieves state-of-the-art performance, i.e. 95.0\% TAR@FAR=0.001\% on IJB-C dataset and 61.98\% Rank1 retrieval rate on the surveillance-to-booking protocol of IJB-S dataset.
Show less

Title: Contributions to Fingerprint Recognition
Creator: Engelsma, Joshua James
Date: 2021
Collection: Electronic Theses & Dissertations
Description: From the early days of the mid to late nineteenth century when scientific research first began to focus on fingerprints, to the present day fingerprint recognition systems we find deployed on our day to day devices, the science of fingerprint recognition has come a long way. In spite of this progress, there remains challenging problems to be solved. This thesis highlights a few of these problems, and proposes solutions to address them. One area of further research that must be conducted on...
Show moreFrom the early days of the mid to late nineteenth century when scientific research first began to focus on fingerprints, to the present day fingerprint recognition systems we find deployed on our day to day devices, the science of fingerprint recognition has come a long way. In spite of this progress, there remains challenging problems to be solved. This thesis highlights a few of these problems, and proposes solutions to address them. One area of further research that must be conducted on fingerprint recognition systems is that of robust, operational evaluations. In chapter two of this thesis, we show how the current practices of using calibration patterns to evaluate fingerprint readers are limited. We then propose a realistic fake finger called the Universal Target. The Universal Target is a realistic, 3D, fake finger (or phantom) which can be imaged by all major types of fingerprint sensing technologies. We show the entire manufacturing (molding and casting) process for fabricating the Universal Targets. Then, we show a series of evaluations which demonstrate how the Universal Targets can be used to operationally evaluate current commercial fingerprint readers. Our Universal Target is a significant step forward in enabling more realistic, standardized evaluations of fingerprint readers. In our third chapter, we shift gears from improving the evaluation standards of fingerprint readers to instead focus on the security of fingerprint readers. In particular, we turn our attention towards detecting fake fingerprint (spoof) attacks. To do so, we open source a fingerprint reader (built from low-cost ubiquitous components), called RaspiReader. RaspiReader is a high-resolution fingerprint reader customized with both direct-view imaging and FTIR imaging in order to better detect fingerprint spoofs. We show through a number of experiments that RaspiReader enables state-of-the-art fingerprint spoof detection accuracy. We also demonstrate that RaspiReader enables better generalization to what are known as "unseen attacks" (those attacks which were not seen during training of the spoof detector). Finally, we show that fingerprints captured by RaspiReader are completely compatible with images captured by legacy fingerprint readers for matching.In chapter four, we move on to propose a major improvement to the fingerprint feature extraction and matching sub-modules of fingerprint recognition systems. In particular, we propose a deep network, called DeepPrint, to extract a 200 byte fixed-length fingerprint representation. While prevailing fingerprint matchers primarily utilize minutiae points and expensive graph matching algorithms for comparison, two DeepPrint representations can be compared with only 192 multiplications and 191 additions. This is extremely useful for large scale search where potentially billions of pairwise fingerprint comparisons must be made. The DeepPrint representation also enables practical encrypted matching using a fully homomorphic encryption scheme. This enables better protection of the fingerprint templates which are stored in the database. While discriminative fixed-length representations are available for both face and iris recognition, such a representation has eluded fingerprint recognition. This chapter aims to fill that void.Finally, we conclude our thesis by working to extend fingerprint recognition to all ages. While current fingerprint recognition systems are being used by billions of teenagers and adults around the world, the youngest people among us remain disenfranchised. In particular, modern day fingerprint recognition systems do not work well on infants and young children. In this penultimate chapter, we aim to rectify this major shortcoming. To that end, we prototype a high-resolution (1900 ppi) infant fingerprint reader. Then, we track and fingerprint 315 infants (under the age of 3 months at enrollment) at the Dayalbagh Children's Hospital in Agra India over the course of 1 year (4 different sessions). To match the infant fingerprints, we develop our own high-resolution infant fingerprint matcher. Our experimental results demonstrate significant promise for the extension of fingerprint recognition to all ages. This work has the potential for major global good as all young infants and children could be given a verifiable digital identity for better vaccination tracking as a child and for government benefits and assistance as an adult. In summary, this thesis makes major contributions to the entire end-to-end fingerprint recognition system and extends its use case to all ages.
Show less

Title: Towards Robust and Secure Face Recognition : Defense Against Physical and Digital Attacks
Creator: Deb, Debayan
Date: 2021
Collection: Electronic Theses & Dissertations
Description: The accuracy, usability, and touchless acquisition of state-of-the-art automated face recognition systems (AFR) have led to their ubiquitous adoption in a plethora of domains, including mobile phone unlock, access control systems, and payment services. Despite impressive recognition performance, prevailing AFR systems remain vulnerable to the growing threat of face attacks which can be launched in both physical and digital domains. Face attacks can be broadly classified into three attack...
Show moreThe accuracy, usability, and touchless acquisition of state-of-the-art automated face recognition systems (AFR) have led to their ubiquitous adoption in a plethora of domains, including mobile phone unlock, access control systems, and payment services. Despite impressive recognition performance, prevailing AFR systems remain vulnerable to the growing threat of face attacks which can be launched in both physical and digital domains. Face attacks can be broadly classified into three attack categories: (i) Spoof attacks: artifacts in the physical domain (e.g., 3D masks, eye glasses, replaying videos), (ii) Adversarial attacks: imperceptible noises added to probes for evading AFR systems, and (iii) Digital manipulation attacks: entirely or partially modified photo-realistic faces using generative models. Each of these categories is composed of different attack types. For example, each spoof medium, e.g., 3D mask and makeup, constitutes one attack type. Likewise, in adversarial and digital manipulation attacks, each attack model, designed by unique objectives and losses, may be considered as one attack type. Thus, the attack categories and types form a 2-layer tree structure encompassing the diverse attacks. Such a tree will inevitably grow in the future. Given the growing dissemination of ``fake news” and "deepfakes", the research community and social media platforms alike are pushing towards generalizable defense against continuously evolving and sophisticated face attacks. In this dissertation, we first propose a set of defense methods that achieve state-of-the-art performance in detecting attack types within individual attack categories, both physical (e.g., face spoofs) and digital (e.g., adversarial faces and digital manipulation), then introduce a method for simultaneously safeguarding against each attack.First, in an effort to impart generalizability and interpretability to face spoof detection systems, we propose a new face anti-spoofing framework specifically designed to detect unknown spoof types, namely, Self-Supervised Regional Fully Convolutional Network (SSR-FCN), that is trained to learn local discriminative cues from a face image in a self-supervised manner. The proposed framework improves generalizability while maintaining the computational efficiency of holistic face anti-spoofing approaches (< 4 ms on a Nvidia GTX 1080Ti GPU). The proposed method is also interpretable since it localizes which parts of the face are labeled as spoofs. Experimental results show that SSR-FCN can achieve True Detection Rate (TDR) = 65% @ 2.0% False Detection Rate (FDR) when evaluated on a dataset comprising of 13 different spoof types under unknown attacks while achieving competitive performances under standard benchmark face anti-spoofing datasets (Oulu-NPU, CASIA-MFSD, and Replay-Attack).Next, we address the problem of defending against adversarial attacks. We first propose, AdvFaces, an automated adversarial face synthesis method that learns to generate minimal perturbations in the salient facial regions. Once AdvFaces is trained, it can automatically evade state-of-the-art face matchers with attack success rates as high as 97.22% and 24.30% at 0.1% FAR for obfuscation and impersonation attacks, respectively. We then propose a new self-supervised adversarial defense framework, namely FaceGuard, that can automatically detect, localize, and purify a wide variety of adversarial faces without utilizing pre-computed adversarial training samples. FaceGuard automatically synthesizes diverse adversarial faces, enabling a classifier to learn to distinguish them from bona fide faces. Concurrently, a purifier attempts to remove the adversarial perturbations in the image space. FaceGuard can achieve 99.81%, 98.73%, and 99.35% detection accuracies on LFW, CelebA, and FFHQ, respectively, on six unseen adversarial attack types.Finally, we take the first steps towards safeguarding AFR systems against face attacks in both physical and digital domains. We propose a new unified face attack detection framework, namely UniFAD, which automatically clusters similar attacks and employs a multi-task learning framework to learn salient features to distinguish between bona fides and coherent attack types. The proposed UniFAD can detect face attacks from 25 attack types across all 3 attack categories with TDR = 94.73% @ 0.2% FDR on a large fake face dataset, namely GrandFake. Further, UniFAD can identify whether attacks are adversarial, digitally manipulated, or contain spoof artifacts, with 97.37% classification accuracy.
Show less

Title: Statistical approaches for the analysis, measurement, and modeling of RFID systems
Creator: Wang, Liyan (Software engineer)
Date: 2018
Collection: Electronic Theses & Dissertations
Description: The goal of this thesis is to develop statistical and learning algorithms for the analysis, measurement, and modeling of wireless networking( Radio frequency identification systems).Radio frequency identification (RFID) systems are widely used in logistic, supply chain industry and inventory management. RFID is already in use in multiple industries and for various purposes. The device in your car that lets you zoom by in the fast lane at a tollbooth, while deducting a dollar amount from your...
Show moreThe goal of this thesis is to develop statistical and learning algorithms for the analysis, measurement, and modeling of wireless networking( Radio frequency identification systems).Radio frequency identification (RFID) systems are widely used in logistic, supply chain industry and inventory management. RFID is already in use in multiple industries and for various purposes. The device in your car that lets you zoom by in the fast lane at a tollbooth, while deducting a dollar amount from your account, is an example of RFID technology in everyday use. Mostly, existing RFID systems are primarily used to identify the RFID tags present in a tag population(e.g., tracking a specific tag from a tag population) while identifying some specific tags is a critical operation, it is usually very time consuming and is not desired or necessary in some situations. For instance, if the objective is to determine whether any of the tags are missing(e.g., to detect some items according to a consignment), the first thing to do is to identify all tags’ ID and then compare with the original record to determine if there is any tags are missing. Definitely, the whole process will be very slow if we have a very large tag population. In this thesis, I present novel statistical algorithms to enable fast and new applications in RFID systems. For example, detecting the missing tags in a large tag population with high accuracy while using the existing infrastructure of RFID systems which is already deployed in industry. More pacifically, I present my work on designing statistical algorithms for estimation the number of missing tags in a population of RFID tags, for detecting and identifying the missing tags from a population of RFID tags.The key distinction of my work compared to prior art is that my methods are compliant with EPCGlobal Class 1 Generation 2 (C1G2) RFID standard. It is critical for RFID methods to be compliant with the C1G2 standard since the commercially available of-the- shelf RFID equipment follows the C1G2 standard. A method which does not comply with the C1G2 standard cannot be deployed on the existing installations of RFID systems because it requires custom hardware, which will cost a lot. In an RFID-enabled warehouse, there may be thousands of tagged items that belong to different categories, e.g., different places of origin or different brands [72]. Each tag attached to an item has a unique ID that consists of two fields: a category ID that specifies the category of the attached object, and a member ID that identifies this object within its category. As a manager of the warehouse, one may desire to timely monitor the product stock of each category. If the stock of a category is too high, it may indicate that this product category is not popular, and the seller needs to adjust the marketing strategy (e.g., lowering prices to increase sales). On the contrary, if the stock of a category is too low, the seller should perform stock replenishment as soon as possible. Manual checking is laborious and of low time-efficiency. You cannot imagine how difficult it is for a manager to manually count the number of items in each category that may be stacked together or placed on high shelves. Hence, it is desirable to exploit the RFID technique to quickly obtain the number of tagged items in each category. A multi-category RFID estimation protocol should satisfy three additional requirements. First, it should be standard compliant; otherwise, it will be difficult to be deployed. Second, it should preserve the privacy of tags by not reading their member IDs. Third, it should work with both a single-reader and multiple-reader environments. As the communication range between a tag and a reader is limited, a large population of tags is often covered by multiple readers whose regions often overlap.
Show less

Title: Attribute prediction from near infrared iris and ocular images
Creator: Bobeldyk, Denton
Date: 2019
Collection: Electronic Theses & Dissertations
Description: The iris is the colored portion of the eye surrounding the pupil. Images captured in the visible spectrum make it difficult for the rich texture of brown irides to be discerned; therefore, iris recognition systems typically capture an image in the Near Infrared (NIR) spectrum. The region surrounding the iris, the ocular region, is also captured by the sensor during the imaging process.The focus of this thesis is on developing methods for predicting soft biometric attributes of an individual...
Show moreThe iris is the colored portion of the eye surrounding the pupil. Images captured in the visible spectrum make it difficult for the rich texture of brown irides to be discerned; therefore, iris recognition systems typically capture an image in the Near Infrared (NIR) spectrum. The region surrounding the iris, the ocular region, is also captured by the sensor during the imaging process.The focus of this thesis is on developing methods for predicting soft biometric attributes of an individual based on the iris and ocular components of the eye. In addition to attribute prediction, the effect of covariates on attribute prediction are also studied. Attributes considered in this work include gender, race and eye color. For the gender and race attributes, both the iris and surrounding ocular region are analyzed to determine which region provides the greatest gender cues. A regional analysis reveals that the iris-excluded ocular region provides a greater gender prediction accuracy than the iris-only region. This finding is of great significance as, typically, the iris-excluded ocular region is discarded by the iris recognition system. This research reinforces the need to retain the iris-excluded ocular region for additional processing. For race, it is shown that the iris-only region provides better prediction accuracy. In order to study the stability of the gender and race features, the impact of image blur on attribute prediction was also examined. It is observed that as the level of image blur increases, the race prediction accuracy decays at a much faster rate than that of gender. For eye color, the textual cues presented on the iris stroma are exploited to generate a discriminatory feature vector that is capable of distinguishing between two categories of eye color. The impact of image resolution on attribute prediction was also determined. A convolutional neural network architecture is presented that is capable of attribute prediction using images as small as 5x6, a mere 30 pixels. Experimental results suggest the possibility of deducing soft biometric attributes from low resolution images, thereby underscoring the feasibility of extracting these attributes from poor quality images. Finally, the thesis explores the possibility of harnessing the feature vector used to predict one attribute (e.g., gender) in order to predict a different attribute (e.g. race). The ensuing experiments convey the viability of cross attribute prediction in the context of NIR ocular images.In summary, this thesis provides insight into attribute prediction from NIR ocular images by conducting an extensive set of experiments.
Show less

MSU Libraries

Digital Repository

Enabled Filters

Pages

Pages

MSU Libraries

Digital Repository

You are here

Search results

Pages

Pages