Content-Based Image Retrieval Based on Integrating Region Segmentation and Relevance Feedback

In the research of content-based image retrieval, visual signature based on region was attracted more attention. To get the signature based on region, the crucial step is image segmentation, and reliable image segmentation is also critical to get the image shape description. Unfortunately, it has been demonstrated that accurate image segmentation is still an open problem. So some strategies for dealing with this problem is to reduce dependence on accurate image segmentation for a practical image retrieval system. Due to the semantic gap, there are still many shortcomings for image retrieval system only with the low level visual features. It is desirable for the relevance feedback based on the user participation in image retrieval system. Through the user's feedback, the corresponding high-level semantic will be obtained based on machine learning theory. Based on ``things cognition'' instead of ``things partition'', the high dimension biomimetic information geometry theory have made good results in many research fields. Based on this theory, to ``cognition'' the whole image characteristic through image segmented into main region and margin region, a prototype image retrieval system was made using the extracted features of color and texture. Combined the retrieval results with relevance feedback technology, image feature dimensional reduction was made using the linear discriminant analysis. It reduces semantic gap and the storage of image signatures, as well as improves the retrieval efficiency and performance. Experimental results on a subset of the COREL database showed the effectiveness of our proposed method.