An Adaptive Fuzzy C-means Algorithm And Chance Constrained Support Vector Machine:Applications To Image Segmentation

Image segmentation is a fundamental problem in image analysis and computer vision, im-age segmentation techniques divide an image into a set of disjointed regions which are satisfied certain similarity criterions by which the image is decomposed into meaningful or spatially co-herent regions. Recently, many mathematical models have been applied in image segmentation, such as the methods based on threshold, region, edge and so on. The methods based threshold can get the segmentation by simply calculating and segment fast. The methods based region are more suitable for homogeneous images. The methods based edge is more applicable to the images with light noise. With the development of image segmentation, FCM algorithm and SVM algorithm are applied to image segmentation. FCM algorithm is unsupervised that re-duced human interference. SVM algorithm requires less sample points based on a powerful generalization capacity. These models have been enhanced by some researchers, but there are still some problems need to solve. FCM algorithm is very sensitive to the parameters. SVM al-gorithm ignores the problem of unbalance data and the need of different accuracy. Two modified approaches for solving image segmentation were proposed in this dissertation. The main results are summarized as follows:FCM algorithm has some shortcomings. First, they are usually very sensitive to the pa-rameters which are supposed to be tuned according to noise intensities. Second, in the case of inhomogeneous noises, using a constant parameter for different image regions is obviously un-reasonable and usually leads to an unideal segmentation result. For overcoming these drawback-s, a noise detecting based adaptive fuzzy c-means algorithm (NDFCM) for image segmentation is proposed in the third chapter. Two image filtering methods, playing the roles of denoising and maintaining detail information, respectively, are utilized in the new algorithm. The parameters for balancing these two parts are computed by measuring the variance of gray level values in each neighborhood. Compared to set the parameter values manually, the parameter is calculated with gray feature of image which improves the adaptability greatly. Numerical experiments on both synthetic and real world image data show that the new algorithm is effective and efficient.In many classification tasks, the requirement for classification accuracy is different from one class to another. A classifier needs not only to deal with nonlinear separable data but also noises and outliers. By introducing slack variable ζ≥ 0, this dissertation develops a kernel method based chance constrained support vector machine with a slack term (KSCC-SVM) in the forth chapter. The modified model aims at mining nonlinear separable data with noises. It is resolved based on chance constrained programming. Both soft classification and kernel method are involved in this approach for a better segmentation result. In the image pre-processing stage, fast generalized fuzzy c-means (FGFCM) algorithm is utilized for labeling samples. Then these labeled samples are randomly selected to train KSCC-SVM. Experimental results show the effi-ciency and robustness of the new proposed algorithm.