Unsupervised Color Image Segmentation Methods Based On Graph Cuts And Quaternion Theory

Color image segmentation is often described as the process of dividing an image into different homogeneous regions with special semantics according to some criteria. As the basis and the key technique of image analysis and image understanding, color image segmentation is always the focus and difficult problem in computer vision and image processing. Graph cuts based on Markov random fields (MRF) formulates the image segmentation as a labeling problem or a energy minimization problem, which has been widely used for interactive image segmentation or object segmentation, but there is little research on unsupervised color image segmentation by graph cuts. The number of segments is difficult to determine automatically and the complex texture and color are included in multi-object images, which bring some difficulties to the unsupervised color image segmentation by graph cuts. Some methods published in the literature on color image segmentation directly use gray image segmentation techniques but fail to consider the coupling between color channels. Recently, because quaternion can handle the coupling between channels of color images, it offers a theory to handle color images as a whole. However, there is large space for researchers to study and explore on how to use quaternion theory to improve further the quality of color image segmentation.To address the above problems, this work provides a deeper study, analysis and summarization about graph cuts and quaternion-based methods of representation and extraction of color and texture features. Some methods to extract texture features of color images based on quaternion theory are proposed. By combining the different texture features with graph cuts, region merging and Gaussian pyramid, this work proposes three graph cuts-based color image segmentation methods and studies how to use graph cuts to automatically segment color textured image with high performance. The main contents of this work are described as follows.A new unsupervised color-texture segmentation method based on multi-scale quaternion Gabor filter (MQGF) and graph cuts is proposed. Using MQGF to extract texture features of a given color image can not only handle the coupling between color channels but also get the scale information of the texture features. The fused features composed of the texture feature and the RGB color feature of the image are modeled with a multivariate finite mixture model, and a splitting framework with s-t graph cuts and minimum description length (MDL) principle is used to segment the image efficiently and automatically. Extensive experiments on synthetic color textured images, Berkeley segmentation database and Weizmann database show that the use of MQGF can improve the discrimination power of the texture features, and the combination of graph cuts and MDL principle can obtain the segmentation results of color images automatically and improve the robustness of the proposed method.A new color image segmentation method based on region merging and graph cuts is proposed, which aims to compensate for the defects of region merging with superpixels and graph cuts, respectively. In this method, a criterion based on MRF energy minimization is utilized to merge the superpixels of a given color image, and the merged result is refined by multi-label graph cuts. Moreover, a two-level Gaussian pyramid is used to reduce the computational complexity of the proposed method. The experimental results show that the proposed method can overcome the shortcomings, such as boundary dislocation and over-segmentation brought by region merging, and at the same time improves the optimization performance of graph cuts.A new method, multi-component graph cuts, is proposed for unsupervised color image segmentation, which aims to address the problems of high computational complexity and over-segmentation during inferring process by a-expansion when the number of labels is too large. Multi-component graph cuts handles the regions within a segment as the multiple components of the segment rather than relabeling them with unique labels. Multi-component graph cuts and a two-level Gaussian pyramid are combined, called two-phase multi-component graph cuts, to reduce the effect of noise and improve the efficiency of multi-component graph cuts. Additionally, two new methods to extract texture features of color images are proposed, including the method based on quaternion cut-off window and the method based on MQGF and texton theory. Based on the texture features and multi-component graph cuts, this work designs two color image segmentation methods. The experimental results indicate that the combination of the multiple component strategy and two texture features can not only improve the discrimination power of features, but also improve the optimization efficiency of graph cuts and alleviate the problem of over-segmentation. Moreover, two-phase multi-component graph cuts can improve the segmentation quality of color images.