The Fuzzy-C Means Clustering Algorithm Based On Spatial Information For Image Segmentation

Image segmentation is one of the key steps in image analysis, description and understanding. The purpose of segmentation is to partition a given image to different regions based on the characteristics. The result of image segmentation directly affects the quality of the subsequent images analysis. It's of great importance that how dividing the target area of interest from a complex background quickly and effectively. This thesis mainly focuses on solving problems of the existing image segmentation, such as low accuracy and poor robustness. The main works are as follows:An image segmentation algorithm based on clone and fuzzy clustering with kernel metric and spatial information is proposed. Firstly, immune clone algorithm is used to get initial cluster centers so that the algorithm does not fall into local optimum. Secondly, the algorithm improves the robustness of noise through adding spatial information to the objective function of fuzzy C-means clustering(FCM). At last, the algorithm uses non-Euclidean distance based on kernels metric instead of Euclidean distance of FCM to enhance segmentation accuracy.A spatial fuzzy C-means clustering algorithm based on mean shift is proposed. In the algorithm, we obtain the cluster number and the initial cluster centers by mean shift algorithm using pixels of the image. Then the pixels in FCM are replaced by the weighted image patches and weighted fuzzy factor is introduced into FCM objective function, which includes spatial and intensity distances of the neighboring pixels at the same time. In addition, the membership is smoothed by the neighboring pixels in each iteration step. By the above improvements, the neighboring information of the image is sufficiently used and the segmentation accuracy is higher and more robust to the noisy image.A suppressed fuzzy C-means clustering image segmentation based on rough set is proposed.Local and non-local information are used to obtain the reconstructed image; then fuzzy c-means clustering algorithm is executed on the histogram of the reconstructed image. The number of histogram is 256, which is much smaller than the number of points. It improves the convergence speed that the clustering is performed on this histogram. Fuzzy C-means algorithm with suppressed strategy also improves the convergence speed by increasing the maximum degree of membership value and suppressed the other membership value. The rough set is used to reduce the influences of the points not belonging to this cluster center and improve the segmentation accuracy robustness.