Segmentation Of Image-based Spectral Clustering Method

Clustering technology, as one of automatic access to information technology. At present, have been a wide range of applications in speech recognition, image processing and other fields. And compared to commonly clustering algorithm, spectral clustering algorithm is able to cluster the sample space of arbitrary shape, and have the benefits which can converge to global optimal solution. In recent years, so, it has been the international machine learning research focus in this area. Image segmentation, is an important applied area of cluster analysis, and is a key step from the image processing to image analysis, taking an important position in the image project. Now, image segmentation has been applicatied in the medical, military, weather, trafficand other social production and life. Spectral clustering has a benefit that is a wide range of adaptability for data samples. Cluster analysis based on pixel and its characteristics is a very vailbe idea for image segmentation. Therefore, this article will attempt to applied spectral clustering to image segmentation, and research and analysis some of the issues. In this paper, we will focus on image segmentation based on spectral clustering method. Experimental analysis on from improving spectral clustering algorithm to apply to the image segmentation. Especially focus research on the texture image segmentation, integrated the texture, color, spatial features, at last, present the experimental results and compared. The paper will introduce as follow:(1) Introduced the basic theory of spectral clustering algorithm, and current research status. Considering the limitation that spectral clustering can not be used for large-scale data processing, we will use the method that combined with Nystrom sampling algorithm and clustering algorithm, and compare the raditonal spectral clustering and complexity in time and space. For the Gaussian kernel parameter sensitivity, we propose a new adaptive neighborhood information parameter selection method with neighborhood information, so, avoid adjusting manually. The experiments show that this adaptive method based on Nystrom sampling spectral clustering algorithm can obtain good results in image segmentation experiments,(2) Describes the basic theory of semi-supervised learning, including several semi-supervised algorithm. The spectral clustering algorithm is extended to the form of semi-supervised for its own characteristics, add pairs of constraints as a priori information to improve clustering performance of the algorithm. Compined adaptive spectral clustering with Nystrom sampling theory, forming a semi-supervised adaptive algorithm. The experimental results show that the method of addition of the semi-supervised segmentation information is better than the ordinary method of spectral clustering.(3) For more complex features of texture images, including color, texture, space and other features, the feature space dimension is more higher, so, we will try to use spectral clustering approach for texture image segmentation, and we will choose Gabor filters to extract texture features. Design a group filter containing the four scales and six directions, to extract multi-dimensional texture feature, and then through smooth filter get the more stable texture features. For feature space contained redundant information, proposed using PCA to extract principal components, and finally get the texture features and using spectral clustering algorithm for image segmentation. In the experiment, we used spectral clustering algorithm for image segmentation based on texture and color features respectively, at last, present the experiment results of integriting the color, texture, spatial feature...