Image segmentation is a technique of dividing an image space into a number of meaningful homogeneous regions. Various data clustering techniques have been adapted in solving segmentation problems. In particular, data clustering is often posed as multi-optimization problem so that characteristics of data could be caught by different objectives simultaneously. Traditional multi-optimization methods often require some prior knowledge or assumptions about data, performance is poor if these assumptions do not hold. Limitations with established multi-optimization methods are caused by their inadequacy in handling a large number of objectives. Non-dominated sorting genetic algorithm III (NSGA-III) is proposed to alleviate this issue. However, NSGA-III is inefficient in removing some bad solutions in high-dimensional searching space during evolution. In this article, we propose a variable string length many-objective genetic algorithm(VMOGA) whose framework has evolved from NSGA-III and its encoding strategy, genetic and evolutionary operator have been redesigned. Performance of VMOGA in image segmentation problems is further enhanced by an appropriate selection of objectives. In the end, we conduct unsupervised segmentation by proposed clustering technique on magnetic resonance image(MRI) of human brain. Comparisons with other evolutionary algorithms are presented and dominance of VMOGA has been demonstrated quantitatively. VMOGA is also performed on detection of delamination area caused by fatigue loading in Mode I glass fiber reinforced polymer (GFRP) samples. Results are compared with fast marching algorithm(FMA) and superiority of VMOGA suggests future potential application in fatigue detection. -- Abstract.
Read
