Image Retrieval Based On Attention Mechanism Deep Hashing

In recent years,multimedia data that include texts,images,audios and videos have exploded due to the widespread applications of social softwares and mobile devices.As a result,searching for one image needed by users in such a large image database has become a new challenge of image retrieval.Benefitting from the remarkable performance of the hashing method in terms of computing capability and storage efficiency and also the strong learning ability of deep neural network in image feature representation,the deep hashing method based on deep neural network has achieved unprecedentedly great success in recent years.In essence,deep hashing method employs a well trained deep neural network to project the high-dimensional features extracted by images into the low-dimensional Hamming space and learns a set of compact hash codes.Based on summarizing and analyzing the research status of hashing methods domestic and abroad,this paper combined some advanced hot topics include deep neural networks,attention mechanisms,feature pyramids and dilated convolution to improve the performance of deep hashing method to a certain degree.Its main research contributions are as follows:(1)Proposed a deep hashing framework based on attention-aware feature pyramid.Most existing deep hashing methods merely encode the raw feature of the last layer for hash learning,neither making full use of the useful information in the middle and shallow layers or ignoring local significant information.To this end,this paper proposes a noval deep hashing method,which employs the channel attention and spatial attention mechanisms to explore the local saliency on feature pyramids of different layers constructed in the deep neural network to learn multiple information of receptive fields.Then adopts a multi-scale feature fusion strategy to integrate the feature representations with the local details and the semantic information.At last,train them to learn high-quality hash codes.(2)A deep hashing framework based on the attention-aware pyramid with dilated convolutions is proposed.Frequent upsampling and downsampling operations in the convolutional neural network can loss the spatial structure information of image.To this end,this paper constructs a noval pyramid model with dilated convolutions instead of conventional convolutional layers.This model not only avoids unnecessary upsampling and downsampling operations,but also expands the receptive field of the feature map at each scale and digs into richer feature information.Moreover,the combination of the channel attention and spatial attention mechanism is adopted a new re-integrate learning way,which further explores the local salient information of the image.(3)A simple and efficient image retrieval system is designed and implemented.This system learns the hash codes of the image database based on the deep hashing method proposed in this paper,and then calculates the Hamming distance between the images according to these hash codes,so as to retrieve the semantically similar images easily and intuitively.