An Improved Affinity Propagation Algorithm With Its Application In Image Processing

Clustering analysis is an important human activity. The application of Clusteringis very wide, such as text mining, image processing and other fields. Affinitypropagation algorithm is a new unsupervised clustering algorithm published in2007onthe Science by Frey and Dueck. It doesn’t need to specify the number of clusters, onlyneed to constructs the similarity matrix，then it can automatically determine theappropriate representative point by message transmission mechanism, and allocate theremaining data to the category that its most similar exemplar belongs to, finally obtainsthe maximum sum of the similarity of all data points to its exemplar. In the APalgorithm, the diagonal values (preference) of the similarity matrix are set to the samevalue at first, and it indicates that all the data points have the same potential to be theexemplar. But this default is defective, because when the number around one data ismore, the possibility of the data point to be the exemplar is larger than the number isless. The main works of this paper is as follows:Firstly, based on the density clustering, count up the number of each data pointwithin the neighborhood ofâ'‚, and put forward a kind of method of setting Preference,thus we put forward a kind of improved AP algorithm that reconstruct the similaritymatrix in this paper.Secondly, apply the MSAP algorithm in image segmentation, and put forward amethod that using the gray histogram to extract the core data in original image to bethe clustering data, therefore, while ensuring the data quality, the data size is greatlyreduced, on the basis, a similarity matrix in image segmentation is reconstructed, at thesame time, a new segmentation evaluation criteria is given. The experiment resultsshow that the MSAP algorithm converges much faster and segments better than the APalgorithm.Finally, apply the MSAP algorithm in image clustering, and put forward a blockand weighted color histogram feature extraction method after the nonuniformquantization of color space, on the basis, a image clustering algorithm based MSAP isgiven, this algorithm use MASP to conduct a preliminary clustering at first, afterobtaining some excellent representative points, then randomly select points as theinitial clustering center of K-means algorithm to conduct a second clustering, it notonly greatly reduces the K-means random initial center influence on the results, butalso solved the problem of the clustering number is not accurate. The experimental results show that the proposed algorithm converges faster, and it achieves the bettercluster result.