Abstract: Relevance feedback algorithm has been an important compose in image retrieval，and it’s a hotspot in the research of image retrieval recently.In this paper，the Reinforcement Learning（RL） based relevance feedback algorithm is proposed.According to Q_learning function of RL，a matrix named Q is established，and for every image counterpart one item Q_i（i=1，2，…，n），which is used to record the image’s total value of feedback in current retrieval.Each feedback，calculates new features depending on weighted features method，then，calculates every image’s current total value of feedback base on Q_learning function.The bigger of the value of Q the more relevant with example image.RL acquires the best path by keeping feed back to the environment，which is consistent the idea of relevance feedback acquires the best answer by grouping the retrieval intention of users.The experiment proves，the algorithmis is more superior.