Spectral Clustering Method In Image Segmentation

The 21st century is an era of information explosion; image has become an important means for human to obtain information, present information and transmit information. Image processing technology as a general term for the visual image processing technology, in people's social and economic life is playing an increasingly important role. Image segmentation is the key to image processing connected to image analysis, and the basis for image understanding and recognition, so, Effective means of image segmentation plays an important role in image processing. Spectral clustering algorithm is built on the basis of spectral theory; compared with the traditional clustering algorithms, it has advantages of clustering in sample space of arbitrary shape and conferencing to global optimal solution, ideally suited for many practical problems, image segmentation naturally included. Therefore, in recent years on spectral clustering algorithm in image segmentation research has received extensive attention of many scholars. Currently, spectral clustering algorithm based image segmentation techniques have obtained some good results, however, because this technology is still in its early research stage, so there are still many problems need to be studied and solved.Not enough stability is one of the problems faced by spectral clustering algorithms, but also is an important reason for that the algorithm is hardly widespread application in image segmentation. Among spectral clustering algorithm, for the clustering algorithm using to cluster the feature vector is sensitive to the initial cluster centers, so spectral clustering algorithms usually have the shortcoming that the stability is not enough, which greatly hinder the algorithm been widely used in image segmentation. For this problem, the semi-supervised spectral clustering algorithm combined with Bayesian decision is proposed. The algorithm adjusts the contents of the similarity matrix, by based on Bayesian decision-making distance learning method, and then improves the distribution of the feature vector for clustering, in order to improve the algorithm stability and accuracy. Meanwhile, the algorithm uses the constrained K-means algorithm to cluster the adjusted feature vectors for clustering, to further improve the algorithm stability and accuracy. Thought is simple and easy to realize; in the use of a large number of unlabeled samples implied by the structure of distribution of information, at the same time, to take full advantage of a small sample of marked constraint information and category information. Experiments show that the algorithm is better than the traditional spectral clustering in a significant improvement both on stability and accuracy.Difficult to apply to massive data is the inherent flaw of spectral clustering, is an important reason for limiting that the algorithm achieve widespread application in large-scale image segmentation. Assuming the size of data sets X is n, and then the size of the similarity matrix W among spectral clustering will be n2. Obviously, for large-scale problems, typically, such as image segmentation, the computation and storage is hard to accept, let alone solving the eigenvectors. Solve the problem, neighboring properties based spectral clustering method for large-scale image segmentation is proposed. This approach firstly obtains the smaller model of an image, which also known as sample images by appropriate sized uniform sampling.Then determine clustering labels of the sample image by the spectral clustering using Nystrom method. Finally, use the above clustering labels to estimate the final classification of the original image based on the rule containing the idea of K nearest neighbor and random selection. Thought is simple and easy to realize; using the spectral clustering algorithm as the core segmentation algorithm can make full use of the advantages that spectral clustering algorithm can cluster in sample space of arbitrary shape and conference to global optimal solution. Experiments show that this method can rapidly and effectively to achieve a good large-scale complex image segmentation.