Research On Content-Based Image Retrieval

Image retrieval is widely used in image management, medical diagnostics, security and copyright protection, which can be divided into two part, Text-Based Image Retrieval (TBIR) and Content-Based Image Retrieval (CBIR).TBIR relies on manual image annotation and its result is subjectivity, what’s more, it’s not suitable for large image database because manualing their annotation is very difficult. While CBIR utilizes visual content of image such as color, texture, shape and so on, and then obtains a more objective result. But there are still many issues, now feature extraction and semantic gap are the main problem of CBIR.Since Color Moment is a typical color feature, this paper analyzed the performance of the RGB and Munsell Color Moments. In order to improve retrieval precision, we proposed a new fusion algorithm which is based on both score and rank called NewcombMNZ to combine RGB with Munsell Color Moments. Experimental evidence suggests that modified algorithm not only improves the retrieval accuracy and rank values, but also shows a good robustness to noise.Color Moment only reflects the statistics information in the image, but it cannot describe the spatial information. This paper does some further study on a feature based on color spatial information, that is, Color Correlogram. This feature usually uses the tradition uniform quantization method resulting in a loss of color information, and can not characterize the real color distribution of image, which leads to a high false reject rate. On the base of our analysis on image color distribution, the Gaussian Mixture Molde (GMM) was found suitable for the description of image color distribution. Therefore, we proposed a GMM-based non-uniform color quantization algorithm which shows a more accurate retrieval result, compared with traditional method.The performance of most CBIR systems is inherently constrained by the gap in knowledge and understanding between low-level features, such as color feature, and high-level concepts, named as Semantic Gap. Relevance Feedback (RF) has been considered as an important approach to bridge the semantic gap in CBIR systems. Firstly, we analyzed the convergence performance of a Relevance Feedback system for the Simple Learning and Active Learning, respectively. And then, we applied Active learning, Color Moment and Color Correlogram to CBIR for reducing the Semantic Gap.