Research And Application Of Image Hash Retrieval Based On Global Distance Estimation

The development of an image retrieval system has essential value for society,which can help human beings to understand images and promote the communication and dissemination of civilization and has good application scenarios.A mature multi-data source retrieval experimental application system can help users quickly and accurately find associated images in different data source scenarios.Achieving accurate and automatic retrieval of multi-data source scenes is essential in developing multi-data source retrieval.To solve this problem,the main research work of this paper is as follows.(1)An image hash retrieval method based on global distance estimation is proposed.For the problem of extracting the hash codes with reasonable spatial distribution under the global perspective,we construct data-dependent hash functions based on the List-wise approach.To adapt to different application scenarios,different distance estimation patterns are designed in this paper for single-label hash retrieval and multi-label hash retrieval,respectively.On the one hand,a learnable distance hyperparameter is designed for single-label hash retrieval to estimate the distance relations of the categories in Hamming space.On the other hand,distance triples are designed for multi-label hash retrieval,and distance relations between categories are estimated,which include not only distance relations between classes but also distance relations between classes when they contain the same other categories.In addition,to make hash codes further reduce the quantization loss of transformation from Euclidean space to Hamming space,corresponding quantization loss methods are proposed in this paper to enhance the accuracy of hash code transformation.(2)This paper proposes a hybrid attention model based on a feature pyramid for further optimizing the global distance estimation-based image hash retrieval method proposed.The method can observe the information of the image from the whole to the local a priori so that the network can better distinguish the massive features of natural images,focus on those semantic features and corresponding detailed information for which the image has local representation with distinguishing ability,enhance the feature representation and effectively enhance the image retrieval level.The experimental results show that after introducing the attention model in the backbone network,the MAP of the network at different bits is improved by 0.645(CIFAR-10 dataset)and 0.63(NUS-WIDE dataset)percentage points on average.Based on the above research,the last part of this paper proposes an experimental application test platform for multi-data source retrieval oriented to the problem that the retrieval model can only be limited for a specific image dataset.The test platform is designed and implemented with a technical architecture that separates front-end and back-end,model microservices and data,and can rapidly integrate multi-data source retrieval resources.Based on the high retrieval accuracy feature,the test platform also has a wide application potential.