An Evolutionary Approach To Unsupervised Color Image Segmentation

Abstract

The process to divide or partition a color image into a set of non- overlapping regions (segments) is called color image segmentation. Color image segmentation can be performed through clustering process by con- sidering the features of each pixel as a pattern and a set of pixels, having similar features or characteristics as a cluster ( segment). Generally, the effectiveness of a clustering algorithm depends on the number of clusters (should be known in advance), initialization of the search space and the searching behaviour of the algorithm. In this work, clustering based algorithms are proposed for color image segmentation which not only determine the number of clusters automat- ically, but also generate compact and well separated segments. First, a hybrid genetic algorithm, called Spatial Fuzzy Genetic Algorithm (SFGA) is proposed which incorporate the colour and spatial information to optimize the fuzzy separation and global compactness simultaneously. The Self Organizing Map (SOM) is adopted to find out the number of clusters (segments) automatically. To initialize the SOM network and SFGA to the productive regions, the dominant peaks in the color his- togram of the wavelet transform image are determined. The problem of over-segmentation is handled with a simple pruning technique. The second contribution is the incorporation of objective function i.e. the ratio of multiple cluster’s overlap to the fuzzy separation into genetic algorithm called Dynamic Genetic Algorithm (DGA). DGA is capable to adjust the number of clusters automatically. Finally, the segmenta- tion of color images are performed by Modified Adaptive Differential Evolution Algorithm (MoADE). MoADE has the ability to automat- ically adjust the crossover and mutation parameters according to the underlying distribution. Moreover to reduce the computational cost the MoADE is applied to the superpixel segmented image. An opposition based strategy is adopted to initialize the population to the productive areas in the search space. The effectiveness of the proposed approaches are tested on Berkeley Im- age Segmentation Database and Benchmark (BSD) with comprehen- sive quantitative and qualitative evaluations. The experimental results demonstrate that the proposed image segmentation methods perform better when applied to complex color images.