Research On Efficient Image Retrieval Based On Deep Supervised Hashing

In digital age,facing a huge amount of information on the Internet,it is of great practical significance to quickly and accurately retrieve the target image of interest.For the large-scale image retrieval tasks,deep hashing-based image retrieval method has been widely used because of its small storage space and fast retrieval speed.However,this method still has a series of problems,such as inefficient feature extraction,insufficient utilization of feature correlation and label information utilization,and weak discrimination of neighborhood ranking relationship of hash code.To solve these problems,three deep supervised hashing algorithms are proposed to achieve efficient image retrieval.The main research contents are as follows:(1)To solve the problem of inefficient feature extraction and insufficient utilization of feature correlation in deep hashing-based algorithms,a new method,Sparse Differential Networks and Multi-Supervised Hashing(SDNMSH),which combines sparse differential network with multi-supervised hashing,is proposed and used for image retrieval.This method guides the learning of hash codes through a well-designed sparse differential convolution network and a general supervised hashing function.The sparse differential convolution network is composed of a sparse differential convolution layer and a common convolution layer.Among them,the sparse differential convolution layer can quickly extract rich feature information,thus achieving efficient feature extraction for the entire network.At the same time,a multi-supervised hashing(MSH)function is used to learn discrete hash codes with distinction by fully utilizing pairwise correlation of semantic information and features,thus achieving efficient image retrieval.The experimental results of the image retrieval on three widely used datasets show that the SDNMSH method is superior to the state-of-the-art hashing approaches.(2)In order to further improve the efficiency of feature learning in the deep network,and because of most of the existing deep hash methods having weak ranking relationships,a new Ranking-based Deep Hashing Network(RDHN)is proposed.It combines the deep feature learning module with the hash learning module,which not only has efficient feature learning ability,but also has better ranking relationship.In the feature learning module,a new type of differential convolution is uniquely designed for the first convolution layer of the deep feature learning network to enhance feature learning ability of the network.In the hash learning module,a new objective function is designed to learn distinguished hash ranking information,which enhances the neighborhood ranking ability of hash codes and reduces the quantification loss.The experimental results on three widely used benchmark datasets show that the retrieval performance of the RDHN method is better than the state-of-the-art hashing approaches.(3)For existing deep supervised hashing algorithms used for multi-label image retrieval,which ranking performance of hash codes is ignored and the label category information is not fully utilized.This thesis presents a deep supervised hashing method,Deep Supervised Hashing with Performance-aware Ranking(PRDH).It can effectively perceive and optimize the performance of the model and improve the effect of the multi-label image retrieval.In the deep hash learning module,an optimal ranking loss function is designed to improve the ranking performance of hash codes.At the same time,a spatial partition loss function is added to divide images with different number of shared labels into corresponding Hamming spaces.In order to make full use of label information,a loss function for multi-label classification is also designed,and the prediction label is explicitly used for Hamming distance calculation in the retrieval stage to achieve the supervision and optimization of Hamming distance ranking.A large number of the experimental results of the image retrieval on three multi-label benchmark datasets show that PRDH method outperforms the state-of-the-art hashing approaches.