EM algorithm with improvement for color image segmentation in multiple color spaces

The goal of image segmentation in imaging science is to solve the problem of partitioning an image into smaller disjoint homogeneous regions that share similar attributes. The novel technique of the expectation-maximization (EM) algorithm based on principal component analysis (PCA) with adaptively selecting dominant factors for color image segmentation in color spaces is studied here. And simultaneously, the final segmentation is completed by a simple labeling scheme. Then the comparative study of the refined EM algorithm is done in multiple color spaces. The experimental results from Berkeley segmentation dataset, illustrate that the improved EM algorithm with or without PCA has good segmentation results with fine adaptability in RGB, CIE XYZ, HSV, YCbCr, and YIQ(NTSC) color spaces where the results of test image changes little. Moreover, the optimized EM algorithm without PCA usually has better segmentation than the one with PCA. Nevertheless, these color spaces, i.e. CIE L*a*b*, and h1h2h3, usually produce poor segmentation on the reliability and accuracy of a set of test images by performance analysis with subjective and objective evaluation indicators.