ImageCluster: An Adaptive Heuristic Image Cluster System

With the development of information science and the Internet, information resources in the Internet have become increasingly diverse, and information is explosive growth. Current most web information retrieval content is web pages. With the exponential growth of web multimedia information and content-based multimedia search technology, multimedia information retrieval become more and more prevalent, such as image retrieval, video/audio retrieval.In the traditional image/web page search engine, because of the ambiguity of query words and retrieval mechanism limitation, it is also very difficult for user to find the information they need from the returned results. As a large number of search results in different topics including link and images thumbnail/summary are displayed to users, they need spend a lot of time looking for what they want. Clustering results according to semantic and other feature is considered an effective way to resolve such issues.Under the guidance of multimedia theory, graph theory, information theory and the theory of pattern recognition, this paper proposes an adaptive heuristic image cluster technology and then designs and implements ImageCluster system.We originally propose a rule-based and statistical algorithms FFRS which fuses the image visual feature and image semantic feature. We use image's visual feature to assist its' semantic feature, so we can create a fusion feature that is more representative of the essential characteristics of the image. In this paper, we propose a adaptive three stage clustering methodology based on color-semantic feature (TSCM). The novel method integrated different levels of image features, took advantage of various clustering algorithms to do the clustering and assigned topic keywords for each resulted cluster through the extraction of keywords.finally; we return the cluster results to the user in the form of star burst. In this paper, we detail the working principle of the algorithms, design and implement an image cluster search engine based on B/S.In the experiments, clusters mean square error and users assessment show that the TSCM and FFRS algorithms have less square error and better user experience.