Study On Several Issues Related To Color Image Segmentation

Over the past three decades, image segmentation has been an active area of research inimage analysis and computer vision, and many algorithms have been proposed to solve thisproblem. Image segmentation is an intermediate step in the image analysis and thesegmentation quality directly affects the quality of the subsequent processing. Imagesegmentation remains an important， but hard-to-solve problem. Since it is related toapplication, so that a prior information of the image structure is usually unavailable. Currently,there is no image segmentation algorithm which is suitable for all different kinds of images.So far, researchers are attempting at utilizing knowledge of artificial intelligence toimage segmentation, some methods have achieved good segmentation results. This paper’stheme is image segmentation. Least squares support vector machine (LS-SVM), Gaussianmixture model (GMM) and hidden Markov tree (HMT) model were studied and summarizedin this paper, completed the following work:1. Based on least squares support vector machine (LS-SVM) theory, the color feature isextracted in consideration of human visual attention, and the texture feature is representedaccording to its intensity content together with local texture content. The training samples areselected by using the two-dimensional Arimoto entropy thresholding, finally classified byusing the least squares support vector machine (LS-SVM). Selecting samples in a way whichis not only fast, but also meet the characteristics of human visual perception.2. A new image segmentation algorithm using spatially neighborhood adapted Gaussianmixture model is proposed based on adapted Gaussian mixture model theory. Simulationresults show that the algorithm segmentation results are in line with human visual perceptioncharacteristics, and have better robustness to noise.3. A new image segmentation method using the PDTDFB-HMT model based on localmean direction of the phase and the module of the coefficient is proposed. First transform theimage with PDTDFB, and then extract local mean direction of the phase and the module ofthe coefficient, finally build an HMT model with local mean direction of the phase and themodule.