Hierarchical feature extraction and interactive query for image database management

Advances in modern multimedia technologies have led to huge and ever growing archives of images, audio and video in diverse application areas such as medicine, remote sensing, entertainment, education and on-line information services. Effective image feature extraction and indexing is an important building block for multimedia information management. In this dissertation, we have explored the color and texture feature extraction and representation for compressed and uncompressed images. We also have developed a new interactive retrieval framework considering users relevance feedback.; We have investigated the effect of color quantization schemes on the performance of image retrieval and developed a new set of color descriptors based on octree data structure. The new color descriptors are more efficient in representing color feature than the traditional multi-resolution histograms because it can describe the color feature of images adaptively. We developed a set of filtering methods to facilitate the retrieval process, including filtering by the dominant color, by the color depth, and by the hierarchical color distributions. A combination of these methods provides a prompt access of image in the database.; Image data are often in compressed form for efficient storage and transmission in a visual information management system. The database population and retrieval process can be greatly speeded up by working directly on compressed images or partially decompressed images. We explored the color and texture feature extraction of images compressed by JPEG. Our basic idea is to extract color and texture information from DCT coefficients, thus indexing can be performed at the same time of compression in the database population stage and retrieval can be performed by partially decompressing the images at the retrieval stage.; The concept, the procedure and the tools for interactive image retrieval with multiple seed images are investigated in this research. Specifically, we consider an interactive query process in which the query can be refined so that the meaning of “similarity” defined by a specific user for a particular application can be approached gradually. This idea is implemented by performing adaptive filtering with multiple low level indexing features based on user's feedback. The proposed approach is demonstrated by a testbed with flexible query formation, efficient initial guess and further refinement tools.