Random Walks Based Image Segmentation Method

Image segmentation is the process and technology of assigning the image into anumber of disjointed set of homogeneous and distinctive regions, and then extracting theinteresting objects. The technology of image segmentation determines the final results andquality of analysis, interpretation and recognition. It has been widely used in many fieldssuch as medical, sports, military, intelligent transportation, industrial product testing,computer vision and so on.In recent years, manual intervention image segmentation based on graph theoryattracted widespread attention in research because of its outstanding advantages, RandomWalks is a typical semi-automatic image segmentation algorithm. Random Walks computesrapidly and can receives good segmentation result, which not only has strong resistance tonoise in the image, but also can detect weak (or missing) boundaries and ambiguousunseeded regions, it is an effective image segmentation method and has great achievementsin research.This thesis concentrates on random walks method based image segmentation andpresents some studies concentrated in the following two topics:1) Improved random walks for image segmentation based on snow modelIn order to reduce the adverse effects of noise in the image segmentation results, adata-adaptive anisotropic filtering technique is proposed to remove noise (simply snowmodel). The size, form and direction of the processing window of the snow model isadaptively steered by image local anisotropic features, so it can preserve the originalsignificant structures while suppressing the noises to the largest extent. At the same time,the traditional random walk algorithm only considers grayscale information of image whendefining the weight function, so this may lead to some inaccurate segmentation results. Theproposed method used grayscale information and gradient information to determine theedges weights, which is more robust.2) Improved random walks for image segmentation based on support value filterThis paper proposes a new image segmentation algorithm, which combines therandom walks with the support value filter. Under the theoretical framework of supportvector machines, it used the mapped LS-SVM to deduce the support value filter, with thebasic support value filter, a series of multi-scale support value filters were obtained by filling zeros in it, the salient features of image is obtained by convolving the support valuefilter and the original image (this is called support value transform). At last the weights ofedges of random walk are determined by both the gray value of original image and thesalient features of support value images, improve the accuracy of algorithm greatly.3) Random walks for image segmentation based on FCMIn order to make random walk can apply to real-time requirements, random walks forimage segmentation based on FCM is proposed, FCM is first used to get the seedsinformation for random walks, then a more accurate result is got by random walks.Combined with the snow model, aiming at removing the noise, the improved randomwalks method is proposed, thanks to support value transform, the robustness, location ofweak or missing boundaries and accuracy of algorithm have great improved. Theexperimental and theoretical analysis proves the superiority of the new method, all of thesenew method and new ideas will further perfect random walks method.