Research Of Image Segmentation By Clustering Analysis

Image segmentation is the basis of implementing vision understanding and has been applied in many fields. Because of the variety of image structure, it is still a hard problem to implement a fast and general image segmenter. The main research aspects of image segmentation includes:building valid segmentation model, lowering the complexity of algorithms, improving the ability of anti-noise and generality of algorithms. The clustering analysis is an unsupervised classification method. When lack of prior knowledge, image segmentation can be finished by clustering analysis. However, the clustering based segmentation methods are not perfect because clustering is sensitive to noise and may converges to local optimum so that the segmentation takes long time.In this thesis, the research status and main problems of image segmentation were examined, then clustering algorithms and image segmentation by clustering were researched in depth according to the characteristics of clustering. Through selection and improvement of clustering algorithms, a fast K-means clustering segmentation based on image edges was designed and implemented, which estimated class initial means by a fissive clustering to avoid local optimum. The experiment result showed the proposed algorithm can automatically segment synthesized and natural images fast and acccurately. The main contents of the thesis includes:1. the main clustering algorithms were analyzed from aspects of computation time and clustering effects. Firstly, we revisited the principles of clustering and summarized all kinds of similarity measure in clustering, then gave the characteristics of these similarity measure. Secondly, the general framework of clustering was examined, meanwhile the concrete clustering algorithms were discussed on the complexity and accuracy of clustering. Finally, it was pointed that it was an effective way to design a valid and practical segmentation algorithm by selecting a suitable clustering algorithm and optimal similarity measure, according to the application background of image segmentation issues.2. the edges are important features of images. It was proposed to sample image edges as clustering samples to reduce greatly the number of samples, accelerating segmentation. Therefore, we researched the principles of edge detection and experimented the main detectors, showing the advantages and disadvantages of these detectors. And these detectors still can be used in color images using the generalized form of image gradient with Euclidean distance measure.3. with edge detection and K-means clustering, a fast image segmentation algorithm was proposed. Firstly, we analysed the problems of K-means clustering and Fuzzy C-means clustering in image segmentation, and discussed the impact of color space on clustering partition, with emphasis in Lab space used by the proposed algorithm. Then, it was proposed to estmate the initial class means using a fissive clustering and preprocess the sample set by median filtering, considering accuracy and noise problems with segmentation algorithms. Clustering algorithms are liable to get trapped into local minima so that undersegmentation happens. By the fissive clustering, the sample set can be partitioned sparsely to avoid converging to overlapped partition and local optimum. The median filtering represents every sample by the median value of neighborhood to enhance similarity space and reduce the impact of independent noise samples. Finally, the detailed steps of proposed algorithm were given and examined by experiments. The expriment result showed the validity and suitability of the proposed algorithm to synthesized and natural images.