科研之友
微信
新浪微博
Facebook
分享链接
摘要
A novel image retrieval algorithm based on k-neighbor semi-supervised affinity propagation algorithm is proposed in this paper. Aiming at the disadvantages of the widespread popularity k-means, such as choosing the initial exemplars randomly, being weak at speed, low efficiency and so on, it is gradually replaced by affinity propagation. However, large-scale image retrieval, the sample points of the local features of image is very large. In the situation, if we apply affinity propagation algorithm to the image clustering directly, there is no doubt about the low speed and out of memory. According to these problems, we combine the k-neighbor consistent and semi-supervised clustering to add to constraints so that the dimension of similarity matrix can be rebuilt from the part distribution of dataset. This method can reduce the size and computational complexity. The messages during affinity propagation algorithm can be transformed in the partial range. Firstly, we extract the integrated features based on color, shape and texture from each image in the image database, and calculate the distance pairs of image features, then establish a similarity matrix. Secondly, k-neighbor semi-supervised affinity propagation clustering algorithm clusters the image in the image database and creates the index. At last, according to the similarity criterion, similarity retrieval in the cluster is carried out. The experiment results show that, compared with the original algorithm, the speed of affinity propagation is faster, precision and recall are more accurate, aiming at the size of dataset.
