Query-free Clothing Retrieval Via Implicit Relevance Feedback

Image retrieval is an important research direction of computer vision,and has wide applications such as e-commerce and public security.In recent years,with the development of deep learning,content-based image retrieval has made significant breakthroughs.This kind of “search by image” method has a premise of a query image.However,many applications cannot meet the above preconditions,such as seeing an interesting piece of clothing on the street or on television show but cannot photograph it.We define this type of problem as “query-free image retrieval”,i.e.the target image to be retrieved exists only in people's mind.This paper takes clothing retrieval as a specific application scenario,and explores the query-free image retrieval problem through implicit relevance feedback.Specifically,the implicit relevance feedback method is that the user only needs to click one image similar to the query image in his or her mind in the pictures recommended by the retrieval system at a time.The system will make the next round of recommendations based on user's clicks and iteratively find the target through this kind of interaction.The most crucial part of this process is how the system makes the next round of recommendations based on user's clicks,which is also the focus of this article.Based on the original research on this issue,we propose a Bayesian-based feature re-weighting algorithm.Different from solely modeling the target image,the Bayesianbased feature re-weighting algorithm proposed in this paper models both the target image and the image features.Through the Bayesian-based algorithm constructed by the two variables as condition variables mutually,the posterior probability of target images and multiple features of images can be updated simultaneously according to the single click of each round of the user.We conduct experiments on this algorithm and various contrast algorithms under the same domain and cross-domain search for 310 real users in the online search system and obtain effective results which lead to massive analysis.Experimental analysis shows that the proposed algorithm can simulate the user's decision-making process.Compared with original algorithms,the proposed algorithm can significantly improve the success rate and shorten the iteration of retrieval completion.Inspired by applications of reinforcement learning,we believe that the query-free image retrieval system is also qualified for learning how to recommend images by itself.To make a new exploration of how to re-weight,we propose a feature re-weighting algorithm based on deep reinforcement learning in the framework of Bayesian-based algorithm.The specific feature re-weighting algorithm is replaced by deep deterministic policy gradient(DDPG)algorithm,which can select actions in continuous space.We made reasonable customization of state,action and reward of DDPG algorithm,and simulate the ideal user to produce a lot of data for training the model.We make a variety of contrast experiments on the ideal user and conduct analysis according to the results.The results show that the proposed algorithm has a significant improvement compared with original ones.However,compared with the Bayesian-based feature re-weighting algorithm,there is a slight gap between them.We conduct a profound analysis of the causes of the gap and look forward to future work.