Research On Deep Hash Image Retrieval Method Based On Attention Mechanism

With the advent of the era of digitalisation and intelligence,more and more images and videos are presented in people’s sight.How to retrieve the target image in the massive image database more efficiently and quickly has become an urgent problem to be solved.The retrieval method based on deep hashing combines deep learning with feature hashing,which has the characteristic of fast speed and great advantages in image feature storage and retrieval.However,when obtaining the feature information of an image through this method,some redundant irrelevant information will be noticed by the network model,which is not conducive to accurately classifying the image,thus affecting the retrieval accuracy.In response to this problem,this paper studies an attention module for cross-dimensional interaction.This attention module consists of three parallel branches,one branch is used to capture the correlation between H and W in the spatial dimension,and the other two branches are used to capture the interaction relationship between channel(C)and spatial(H,W)dimensions.This attention module can be applied in image retrieval tasks to improve the performance of convolutional neural networks and learn more feature information that is conducive to retrieval.This paper conducts comparative experiments on the CIFAR-10 and NUS-WIDE datasets,verifying that this attention module can effectively improve retrieval accuracy.In deep hash image retrieval method,high-dimensional image features trained by the convolutional neural networks are usually converted into binary hash values,and then the retrieval task is completed.However,there is a quantization error in the hash encoding process,resulting in weak distinguishability of the obtained hash values.In order to further reduce the quantitative error in hash encoding,this paper designs a deep hash image retrieval method based on greedy strategy.The VGG16 network is used as the feature extractor to extract the key feature information of the image in combination with the above attention module.The greedy strategy is used to optimize the binary hash code,so that the model can directly generate the binary code in the training phase,retain the discreteness of the hash code,and add a penalty term to the loss function to solve the problem of gradient disappearance.This paper conducts comparative experiments on the CIFAR-10 and NUS-WIDE datasets,verifying that this deep hash method has higher retrieval accuracy in image retrieval tasks compared to other classic deep hash methods.Based on the above research content,this paper designs and implements an efficient and simple image retrieval system.The system encodes the image data through the deep hash image retrieval method designed in this paper,and finally returns the retrieved image according to the actual needs of the user.The system has good stability and fast retrieval speed,which realizes the user’s demand for "image search".