Affine Spread Application Of Clustering Algorithm In Image Retrieval

With the development of information technology, digital image information features informative straight forward and understandable and is involved and shared in all walks of life more and more widely. Of course, when the amount of digital image information is increasing, it is a challenging job how to realize digital image retrieval fast and effectively. The traditional method is that the feature of query image and database are matched according to similarity by sequence method. The image which has the maximal similarity is the final result. However, the more images in the database, the larger the feature database scale. The problem of sequential retrieval becomes the bottleneck of the image retrieval system. If image feature database is treated by making use of one clustering algorithm according to similarity, we can set up the index scheme to retrieve then. Clearly, this method narrows the retrieval scope and plays a key role in finding out image accurately and effectively. Meanwhile, the time of matching the independent images is shortened. Under this background, this essay emphasis on the study of feature vector in clusters.First, this paper introduced the status of and several key factors of image retrieval. This discussion focused on the high-dimensional indexing technique base on visual keywords. For the characteristics of high-dimensional data in the visual keywords, we studied the affinity propagation algorithm which has very good effect to high-dimensional data. We also discussed the advantage and research status of affinity propagation algorithm being used to cluster visual keywords. Second, through the analysis on self-problems of affinity propagation clustering algorithm, when k-means algorithm is replaced by affinity propagation, there has following several problems:(1) the computational complexity of the large-scale data sets using affinity propagation algorithm;(2) the optimal number of clustering with affinity propagation. If the affinity propagation algorithm can be applied the composition of visual keyword, it will probably constitute the optimal vocabulary tree based on visual keywords. And it is quite helpful to the performance of image retrieval. Considering accuracy and speed, there are two methods to solve the above problems:(1) using semi-supervised affinity propagation based on k-nearest-neighbor model to deal with large-scale data;(2)choosing the optimal the number of clustering based on the level division of affinity propagation. Finally, to prove the validity of these algorithms, many simulation examples are given under several image databases, such as Corel, LabelMe and Caltech101and so on. After that, we begin to retrieve on the basis of the results of clustering. The retrieval results of the running time, precision and recall of the original affinity propagation and the improved affinity propagation can be obtained by the retrieving interfaces. The comparison on above results shows that the improve affinity propagation decreased the computation cost and improved the retrieval speed as far as possible, under the guidance of guaranteeing retrieval accuracy. Meanwhile, the improved algorithm is more suitable to the content-based of large image databases.