Research On Some Key Techniques Of Content Based Image Retrieval

Demand for effectively managing the huge amount of image data becomes more and more urgent,because traditional retrieval methods based on keyword and description text are not competent for it while resources of images and videos becomes more and more abundant in modern information era. Content based image and video retrieval techniques have emerged as the times require,which combine the multi-disciplinary such as image understanding,computer vision,database technique and artificial intelligence etc. It has many advanced aspects such as objective and flexible description,highly automatic input process and wide application areas etc,which has been paid more and more attentions and developed rapidly. In this dissertation,content based retrieval techniques for a static image and dynamic video are deeply researched,and some valuable results are achieved based on a thorough investigation of the current techniques.The research works of this dissertation for static image retrieval mainly include following parts. (1) The basic elements and evaluation criteria of image content representation are summarized and a retrieval method for DCT compressed imaged based on codebook is proposed,which could combine multiple features. (2) A concept and relative method for content based retrieval technique based on spatial layout of image regions are proposed. It includes following parts. An image segmentation method is proposed,which fuses color and texture features. A similarity measure combining multiple features and layout is also presented. Two attributed graph match methods based on local searching and genetic algorithm are proposed to meet the occasions of different applications. (3) A static image retrieval testing system is designed and implemented. It includes following subparts. A hierarchy image content description model is proposed. A multi-feature indexing tree structure using Kohen neural network clustering technique is proposed. A relevance feedback mechanism is also proposed. The proposed all algorithms for content based static image retrieval are applied to a database with about 1000 images,and the experimental results are prospective and satisfied.The research works of this dissertation for dynamic video retrieval mainly include following parts. (1) A shot detection and key frame extraction algorithm based on codebook and region matching is proposed,and a representation model of shot combining multiple features is also built. (2) Methods for detecting faces and extracting texts from videos are presented,and an algorithm for judging camera motion type is also implemented. (3) Based on hyper-graph clustering model,detection and description algorithm of a story unit is presented,and an algorithm frame for judging the type of story units is also proposed. (4) A video retrieval testing system is devised and implemented. It includes following subparts. A multi-granularity description model of video is proposed. Similarity measures for comparing videos at different granularity levels are also developed. The proposed allalgorithms for content based video retrieval are applied to video sequences of famous story films for automatic processing,which consist of detection shots,story units and so on,and the experimental results are prospective and satisfied.