Deep Semantic Based Hashing And Relevance Feedback In Image Retrieval

The retrieval performance of a content-based image retrieval system crucially depends on the learning of effective feature representations and similarity measures.Despite research efforts for several decades,it still remains at the exploratory and research stage.The “semantic gap” issue which exists the low-level image pixels captured by machines and high-level semantic concepts perceived by human has hindered the development of content-based image retrieval.With the success of deep learning in computer vision and other applications,it may solve the "semantic gap" issue and learn effective feature representations.In this paper,by analyzing the research status and problems of CBIR,we improved the network structure based on the Alex Net network and added a hashing layer.An end-to-end deep semantic based hashing(DSBH)was proposed to improve the system's ability to learn images and search quality.DSBH embeds more semantic information into the hash layer,which makes the hashing binary code contain more semantic information.The experimental results show that DSBH improves the retrieval efficiency and accuracy of the system,and solves the "semantic gap" problem to a certain extent.In order to further solve the “semantic gap” issue,we proposes a new type ofRelevance Feedback(NRF)mechanism to retrain deep CNN through feedback information from multiple users.This new method optimizes the search results from the global rather than the traditional local optimization.From our empirical studies,by combining with deep semantic based hashing called DSBH-NRF,DSBH-NRF improves the retrieval accuracy of the system and solves the "semantic gap" problem effectively.