Vocabulary Tree-based Affine Propagation Clustering Algorithm To Generate Research

The development of network multimedia brings abroad attention and research for massive image information management and retrieval technique. There is an urgent requirement for efficient retrieval system for large-scale image database. When scaling up to very large database of images, the retrieval tasks generally need for an efficient indexing scheme to avoid scanning through the entire database sequentially. In this paper, a high-dimensional indexing structure that scales efficiently to a large number of images is presented.Through the research and analysis of the structure and key technologies of content-based image retrieval system, the thesis studies the image database indexing scheme and constructs a vocabulary tree structure. Images are represented as sets of PCA-SIFT local features, each of which is a vector describing a small patch of the image. The algorithm applied incremental clustering techniques on large groups of local features from many images, formed a vocabulary of visual words by partitioning the feature space. Each visual word in the tree is associated with an inverted file that statistics the visual word frequency of each image. Thus, each image is represented simply as a set of these discrete visual words. The visual words are used in hierarchical TF-IDF scoring of the relevance of an image to the query. The local region descriptors are hierarchically quantized in a vocabulary tree. When the number of visual words is large, instead of having to scan through the entire vocabulary to find matching images, a tree structure allows vocabulary lookups in a short time.In this thesis, I describe the concept, structure and growth process of the vocabulary tree which based on K-means clustering. To validate the feasibility and superiority of image retrieval with vocabulary tree, I give an experimental evaluation for it. Further investigation shows that the weaknesses of k-means clustering would result in deviation of the vocabulary tree structure. Based on survey of related papers and further study on the pattern recognition of cluster analysis theory, I develop an improved vocabulary tree structure which based on affinity propagation clustering algorithm. There were three datasets used to test the tasks:the Corel dataset, the LabelMe dataset and the Caltech-101 dataset. The experiments evaluated this new build method for vocabulary tree offers not only incrementally computes sets of vocabulary tree quickly, but gains in retrieval accuracy as well.