Design And Implementation Of An Image Retrieval System Based On Reinforcement Learning With Multiple Motivations

The image retrieval technology enlarges the areas of traditional information retrieval. Inearlier times, the research on content-based image retrieval has been proposed, manyresearchers have proposed constructive approaches, and some of them have been intoapplication. This technique has been a hot research area and a lot of companies havedeveloped the application systems.In the content-based image retrieval, users do not search with key words, but search withimages. The key point is getting some key information from the images and match them in theimage database, then find the results which most meet the users’ requirement. Manycompanies such as Google and Baidu have developed content-based image retrieval systems,but because of the difference between machine and human, so the existing content-basedimage retrieval systems are not perfect. So the existing content-based image retrieval systemsare only assistant ones.In this paper, I mainly researched the major parts of the image search engine. This thesisdescribed the design and implementation of an image retrieval system based on reinforcementlearning with multiple motivations. Furthermore I built an experimental verification system,and the experimental results illustrate the effective of the algorithm.The image retrieval technology enlarges the areas of traditional information retrieval. Inearlier times, the research on content-based image retrieval has been proposed, manyresearchers has proposed constructive approaches, and some of them have been intoapplication. This technique has been a hot research area and a lot of companies havedeveloped the application systems.In the content-based image retrieval, users do not search with key words, but search withimages. The key point is getting some key information from the images and match them in theimage database, then find the results which most meet the users’ requirement. Manycompanies such as Google and Baidu have developed content-based image retrieval systems, but because of the difference between machine and human, so the existing content-basedimage retrieval systems are not perfect. So the existing content-based image retrieval systemsare only assistant ones.In this paper, we proposed a content-based image retrieval approach based on multiplemotivation reinforcement learning. In this approach, for each group of train data, we set aninitial value for the Agent randomly, then making retrieval to the image, the Agent calculatescurrent states (the feature vector of the image), then select an action (select an image from theimage database) to get a new image according to the action selecting strategies. If theresulting image is similar to the former one, then it means the result is relative good, and theAgent get a positive reward, and update the Q value matrix, loop until the reward can be notincreased, then we believe that the Q value matrix is good.