Research Of Image Segmentation Algorithm Based On Clustering Analysis

Clustering analysis is an important technology in data mining, and it has been widelyused in areas such as statistics, image processing, medical diagnosis, information retrieval,biology and machine learning. Obviously, the problem of image segmentation is coincidentwith the problem of clustering. Image segmentation plays a fundamental role in computervision as a requisite step in such tasks as object detection, classification, and tracking. Incomputer vision, a lot of clustering algorithms are not proper to be applied if the image dataset is too big. Because the data of digital image processing in computer vision is often greatand the complexity of the image itself is unpredictable, especially for the treatment of thenatural color image sequence which is very large, the processing of image segmentationoften takes more time and lacks of precision or it is unrealistic. Therefore, how to realize thereal-time automatic quick image segmentation is still extremely important and the problemsare yet to be solved effectively.According to the problems proposed above, this dissertation studies the key problemsof many clustering algorithms existing large storage and computational complexity and hardto be applied to large scale images segmentation, and obtains the following innovativecontributions:(1) In order to solve the problem of the traditional Ncut algorithm existing largestorage and computational complexity, a new color image segmentation method combiningtwice used watershed and Ncut of improved the weight matrix algorithm is presented in thispaper. The image clustering uses the segmented regions, instead of the image pixels, thenew method can effectively reduce the computational complexity of traditional Ncutmethod by using secondary watershed algorithm. The new weight matrix also has certainself-adaptability.(2) According to the problems of the affinity propagation(AP) clustering algorithmexisting huge storage and computational complexity and hard to be used in image datareal-time processing, a new color image segmentation algorithm is proposed based on meanshift(MS) and affinity propagation algorithm named MSAP. The proposed methodpreprocesses an input image by MS algorithm. The numbers of segmented regions, instead of the numbers of image pixels, are considered as the input data scale of AP algorithm. Theaverage of the color vectors in each region is calculated and considered as an input datapoint of AP algorithm. Distances between data points are regards as similarity measureindex, and then the AP algorithm is applied to perform globally optimized clustering andsegmentation based on similarity matrix.(3) Hierarchical clustering(HC) algorithm can obtain good clustering results, but itneeds large storage and computational complexity for large image processing. A new colorimage segmentation algorithm based on mean shift and hierarchical clustering algorithmnamed MSHC is presented in this paper. MSHC algorithm preprocesses an input image byMS algorithm to form segmented regions that preserve the desirable discontinuitycharacteristics of image. The number of segmented regions, instead of the number of imagepixels, is considered as the input data scale of HC algorithm. The proximity between eachcluster is calculated to form the proximity matrix, and then ward algorithm is employed toobtain the final segmentation results. MSHC algorithm is employed on color image andmedical image segmentation.(4) In order to obtain good and robust clustering results, a new clustering algorithmcalled clustering by data competition is proposed for large image segmentation processing.The proposed algorithm the clustering exemplars and clustering members according to thedata energy and determines the best exemplars according to the data competition. Then theproposed algorithm is combined effectively with the mean shift algorithm for large scalecolor image segmentations. Furthermore, the proposed algorithm has high segmentationefficiency, and gets a better image segmentation performance as well.(5) In order to solve the problem of the traditional spectral clustering existing hugestorage and computational complexity and hard to be used in image data real-timeprocessing, the cosine similarity is introduced into the process of image spectralsegmentation. A fast image segmentation algorithm based on spectral clustering of Nystr mapproximation and cosine similarity is presented. That the cosine similarity is calculated asthe weight matrix of spectral clustering avoid to calculating the exponential operation andparameter setting and reduce the computational cost effectively.