Neural Network-based Image Retrieval Relevance Feedback Mechanism

Recently, the content-based image retrieval (CBIR) technique has been one of top important research domains. It expresses, recognizes and understands the content of image by extracting the visual information contained in the image, then retrieves image according to comparability measure. But it is different that human and the computer understands the comparability about two image, that is, there is a gap between the high level notion of human and low-level visual features of image. So, the image retrieval system can not retrieve satisfactory results. As a kind of human-computer technique, relevance feedback can make up the computer's limited ability of understanding and greatly reduced the gap between the high level notion and low-level visual features by cooperation between human and computer. The system can sharply improve retrieval precision by using relevance feedback.In this paper, we main research the relevance feedback technique in image retrieval. On basis of this, we presented a based-RBFN relevance feedback algorithm in the view of machine learning. The main contributions of this paper are:(1) Profound research about relevance feedback technique. It profound researches the idea, the interactive procedure, and the user's measure model of the relevance feedback. Then, it presents a three-grade relevance feedback method: relevance, fuzzy relevance and no-relevance. This method gives users much convenience when users label the retrieval results.(2) Research about extracting the color feature. It presents a block color histogram algorithm on basis of analyzing color histogram. This method first divides the image into 3X3 no-equality blocks, then respectively extracts color histogram from each block. It combines color and spatial information. Experimental results show that this method can efficiently express the color features of image, its retrieval results are better than the global color histogram.(3) It presents a based-RBFN image retrieval method with three-grade relevance feedback. The users label the first retrieval results: relevance, fuzzy relevance and no-relevance, then return the system. Based on the users' feedback information, the system constructs an RBFN, and the underlying parameters and network structure are optimized using a gradient-descent training strategy.At last, the system retrieves image from the database. Experimental resultsshow that this method is efficient, and can retrieve more satisfactory results inthe less number of iterations.(4) This paper designs a content based image retrieval system with relevancefeedback mechanism as the test bed for retrieval and relevance feedbackalgorithms, which is an experimental frame system.This paper verified the efficient of algorithms presented in the paper by the above work. It makes the relevance feedback technique have more effect on thedomain of image retrieval and image understanding.