The Research Of Relevance Feedback Based On Reinforcement Learning In Image Retrieval

Along with the universalization of computer Internet and various digital equipments, multi-media information management research, in particular image retrieval, attracts more and more attention. At present, the content-based image retrieval technology becomes the mainstream of image retrieval. In order to improves the search results, the relevant feedback technology which realized a good man-machine interaction has been introduced in image retrieval, therefore relevance feedback become an indispensable part of image retrieval.In this paper, on the basis of researching content-based image retrieval, relevance feedback algorithm and reinforcement learning, integrating the Q learning algorithm into the relevant feedback, a relevance feedback algorithm based on Q learning is presented. The experimental results are intending and good. In addition, the paper aslo improved the existing IRRL model (Image Retrieval Reinforcement Learning model) alogrithm, comparing with the traditional algorithm the new algorithm has better efficiency.The main work of this paper contains:1. The existing visualize feature extraction algorithms, in particularly feature of color and shape extraction algorithms, have been analyzed, and a better feature extraction algorithm has been proposed derived from specific experiments.2. Through the study of reinforcement learning theory and Q learning algorithm, the thesis is to find the integration points with relevance feedback algorithm. Choice Q matrix records each image's cumulative reward value; through the Q learning algorithm the greater Q value of the image is more relevant according to the final Q matrix.3. Amendment the existing IRRL model, it's to extract the optimal strategy for each feedback by choosing the middle value image for each type of images, in accordance with the latest round feedback value of user to determine the choice of the relevant feedback algorithm.4. After acquired the strategy, when put up the image retrieval in accordance with the strategy, there are a problem about the combination between differences feedback algorithms, the thesis access to a more good feedback result by the combination of the weighted feature, query optimization and Bayesian classifier.This thesis makes the relevance feedback can achieve better HCI by introducing the reinforcment learning principles related to the image retrieval feedback technology, retrieval system can better meet the user's search intention.