Study Of Image Segmentation Based On Affinity Propagation Clustering

Image segmentation is an important technology of image processing, and has attracted great attention. Since the advent of image segmentation technology, thousands of different segmentation methods have emerged, but still no generic one. Images to be segmented in different areas are varied, requiring different yet appropriate segmentation methods. With the continuous development of clustering technology, image segmentation method based on clustering technology with good performance keeps popping up.Affinity Propagation Clustering is a new clustering algorithm, and is famous for accuracy, stability and clustering without preset cluster number and initial cluster centers. Studies using affinity propagation clustering for image segmentation is currently relatively short. Affinity propagation clustering works mostly based on the affinity between data points. So the measure of affinity between pixels directly affects the result of the segmentation when AP is used for image segmentation.Fuzzy connectedness segmentation algorithm has its own independent set of theoretical system, and spawned a lot of algorithms. Fuzzy connectedness is a fuzzy relationship between pixels to pixels. It takes into account the spatial information of pixels, but won’t over segment an object when the object spans very large. It is suitable as a basis for calculating the similarity between pixels and pixels. Fuzzy connectedness algorithm runs fast, performs very well, and has been widely used in medical image segmentation.This paper studies image segmentation based on affinity propagation clustering and fuzzy connectedness. The main innovations are as follows:1) A full fuzzy connectedness algorithm on the basis of fuzzy connectedness theory is proposed. Full fuzzy connectedness algorithm can calculate the fuzzy connectedness and an optimal connection path between all data points based on the input of the affinity matrix. This paper describes the algorithm and gives proof system, convergence and complexity analysis.2) A FCAP algorithm is proposed by combining full fuzzy connectedness algorithm and affinity propagation clustering algorithm. In FCAP, fuzzy connectedness between pixels are calculated based on their spatial relationships and features. Then affinity matrix is computed based on it. Finally AP clustering is used to finish the segmentation.3) A segmentation method of natural color images is proposed by combining Normalized Cut and FCAP. FCAP color image segmentation method pre-segments the image into super pixels using Normalized Cut. Then FCAP is used to complete segmentation according to the location and the average LUV color feature. Normalized Cut can greatly reduce the amount of data.4) A new super pixel texture extraction method is presented. Traditional GLCM computation method based on sliding window has many disadvantages, such as window crossing border, exceeding image range and difficulty to set window size. This paper proposes a new method using super pixels as the window. Then the super pixel texture extraction method is combined with FCAP in applying to remote sensing image segmentation.