Research On Image Retrieval Methods Based On Deep Hashing

With the advent of the multimedia era,multimedia data such as images and videos are growing explosively,and how to retrieve massive image data efficiently has become an important research direction.Deep hash methods combining deep learning and feature hashing have achieved remarkable retrieval performance in coping with the retrieval problem of large-scale image data,but it still faces the challenges of large intra-class variation and small inter-class variation in image sets,and also suffers from the performance impact caused by the noise,visual variation,illumination variation,deformation,background interference and occlusion overlap situation existing in the image.Therefore,in response to the above problems,this thesis mainly focuses on three aspects of the deep hash methods: the feature expression ability,semantic information utilization and training manner.The main research work of this paper is as follows:(1)To address the problems of low learning efficiency,limited feature expression capability,and insufficient utilization of semantic information in existing deep hash methods,this thesis proposes a hashing method combining the attention mechanism and central similarity learning(Attention-based Hashing with Central Similarity Learning,ACSH).First,a set of semantic hash centers is generated for the image set,and the hash codes are learned from the global data distribution effectively through central similarity learning to improve the learning efficiency of the model.Then a spatial attention mechanism is embedded in the feature extraction module to learn the spatial weight distribution of significant regions of the image and apply to the original features to improve the feature representation of deep hash network.Finally,a classification task is added to supervise the learning of important features by the attention mechanism,which can capture more complex semantic information of the image.The experimental results show that compared with the CSQ method,the ACSH method only increases in hash coding time and space complexity,but none of the increases are significant,while the ACSH method improves m AP@1000 by 2.5%,1.8%,and 1.3% on Image Net with different number of code bits,respectively,and improves m AP@5000 by 0.8%,2.9%,3.4% on MS COCO,and 2.0%,2.6%,1.9% on NUS-WIDE with different number of code bits,respectively.(2)To address the problem that most existing deep hash methods treat all samples equally and do not take into account the impact of sample imbalance on the training of deep hash networks in complex scenes,resulting in poor generalization ability of the model in complex scenes,This thesis improves the training manner on the basis of the CSQ method and proposes a deep hash method combining weighted pairwise similarity learning(Deep Hashing with Weighted Pairwise Similarity Learning,DH-WPSL).It first adds a pairwise similarity loss term to the loss function of the CSQ method,in which different weights are assigned to simple sample pairs and hard sample pairs according to the similarity relationship of the image pairs’ hash codes,making the hash network more concerned with the training of hard samples,and it optimizes the deep hash function through central similarity learning and pairwise similarity metric learning jointly,thus improving the generalization ability of the deep hash network model based on central similarity in complex scenarios.The experimental results show that compared with the CSQ method,the DH-WPSL method does not add additional time and space complexity,while the DH-WPSL method improves m AP@5000 by 0.9%,1.1%,2.1% on MS COCO,and 3.1%,3.1%,2.3% on NUS-WIDE with different number of code bits,respectively.(3)In response to the need to focus on contextual relationships in images in complex scenes and the problems of category confusion and obscure categories,this thesis improves the CSQ model from three aspects: feature expression ability,semantic information utilization and the training manner,and proposes a deep hash image retrieval method for complex scenarios(ACS-WPSL)based on the ACSH method and DH-WPSL method,and performs experiments on the two benchmark datasets.The experimental results show that the ACS-WPSL method adds extra space complexity,but the increase is not significant,while the ACS-WPSL method improves the m AP@5000 by 1.4%,3.1%,3.5% on MS COCO,and 3.4%,3.5%,2.4% on NUS-WIDE with different number of code bits,respectively.Finally,this thesis applies the ACS-WPSL model to the image retrieval technology for Flickr image service platform in complex scenarios.The application results show that the ACS-WPSL method takes less time and achieves an average query accuracy of over90%,indicating that it can still achieve relatively good retrieval performance in complex scenarios,proving the feasibility of the ACS-WPSL method in application.