The Research Of Image Retrieval Technique Based On Illumination Invariance And Region Auto-Segmentation

Content-based image retrieval (CBIR) is one of the Content-based retrieval techniques (CBR). It uses color, shape, texture and semantic features in image itself to get the desired images based on the understanding of image. CBIR constructs image indices by extracting image features directly from images. This process can be accomplished by computer automatically, thus avoiding the subjectivity of human description and amount of work. CBIR differs from traditional retrieval methods by integrating image understanding technique. It can provide more effective retrieval means and realize automatic retrieval. CBIR is a complex task, which involves technical field like image understanding and processing, pattern recognition, artificial intelligence, similarity search and data base technique, et al. The advantage of CBIR will certainly make it play an increasingly important role in image retrieval domain, and will have a broad development space.In this paper, we focus on the hot issues in content-based image retrieval, and mainly study several techniques that the image retrieval method based on illumination invariance and region auto-segmentation involved. The main research achievements include:(1) Since color features are too easy to be influenced by illumination changes, we propose a novel illumination invariant color feature descriptor, i.e. the gradient correlogram. The descriptor effectively overcomes the disadvantages of the existing illumination invariant color feature descriptors on the large size of the feature vector, poor robustness to the shifting, scaling and viewpoint of object.(2) We propose a novel RBIR method which based on region auto-segmentation. The method starts with the image segmentation process which is based on the k-means clustering algorithm. Then we extract the color and shape feature in each region as well as the new presented auto region correlogram feature as the integrate features to represent the image on region-level. Finally, we calculate the similarity among different images using our new defined quadratic distance similarity measure (QDSM).(3) Because the "region-to-region" similarity measure method used in RBIR has disadvantages in simplifying the querying interface and reflecting the overall image information, we propose a novel "image-to-image" similarity measure, which is called the quadratic distance similarity measure (QDSM). QDSM incorporates the properties of all the segmented regions so that the information of an image can be fully utilized, and it can be applied to images with different number of segment regions.(4) We introduce relevance feedback to the region-based image retrieval system, and propose a novel weighted average feature fusion method which is based on ISODATA clustering to implement feedback process in RBIR system.