Research On Deep Hashing For Complex Scene Image Retrieval

With the development of information technology and the popularity of online social networking platforms,a large amount of image data has appeared on the Internet,and the fast retrieval of these images is one of the core tasks of Internet-related industries.The hashing retrieval method is an important method for fast image retrieval,which has attracted extensive attention from researchers in recent years because of its good performance.By mapping image data into discrete binary codes with a fixed length,the image hashing retrieval method can achieve fast approximate nearest neighbor retrieval with high efficiency and low storage capacity.However,with the improvement of the software and hardware technology of image acquisition equipment,a large number of complex scene images which contain indiscernible background or multiple entities have been generated while satisfying user needs.Compared with general images,the complex scene image has put forward higher requirements for hashing retrieval methods.For complex scene images,there are two main solutions in the research of hashing retrieval:(1)constructing more efficient feature extractors to extract features containing more image information for generating binary hash codes,so as to improve the representation ability of the hash code;(2)designing more effective objective functions to guide the optimization of the hash code to a better distribution direction,thereby enhance the discrimination ability of the hash code.Based on the above two ideas,we propose Deep Regional Detail-Aware Hashing(DDAH)and Deep Hard Sample Mining Hashing(DHMH),respectively.(1)In terms of feature extractor ideas,to extract more image information,it is necessary to deeply mine the detailed information of the image.Most of the existing deep image hashing methods directly generate hash codes for the entire image,which can easily cause the ignorance of the local details in complex scene images,thus reducing the representation ability of hash codes.To this end,we propose the Deep Regional Detail-Aware Hashing which modifies the network structure so that the network can generate features which approximate hash codes of local regions from an image,and then fuse them into a global hash code.On this basis,we also use the overlapping local regions of the same image as similar pairs,which can alleviate the class imbalance problem inherent in pairwise label-based hashing learning methods on the one hand,and optimize the distribution of hash codes on the other.We verify the effectiveness of the method through detailed experiments on three publicly available datasets.(2)In terms of objective function ideas,a better distribution of hash codes means that complex scene images are easier to distinguish,which requires improving the attention to complex scene images to optimize the distribution of hash codes during the training process.To this end,we try to reduce the weights of simple samples and give higher weights to difficult samples during the training process,which could improve the retrieval ability of the hash learning method for difficult samples.This solution is consistent with the idea of hard sample mining.Therefore,we introduce the hard sample mining function and adapt it to the characteristics of the hash retrieval task(i.e.,the actual similarity of positive and negative pairs is not consistent with the similarity identified at retrieval process)to better meet the requirement of the hash retrieval task.The experimental results on three publicly available datasets show the superiority of mining hard samples.