Deep Hashing Method For Image Retrieval With Explicitly Guided Attention Mechanism

For a long time,with the rapid development of emerging social networking sites and mobile terminal imaging technology,visual data such as photos and videos have shown an explosive growth trend.These visual data contain a huge amount of information,which make it important to explore an efficient and effective retrieval algorithms for visual data.This problem faces these challenges:accurate nearest neighbors search methods have high requirements on time and computing power;while the visual data has problems of storage and computational efficiency,and difficulty in transferring from visual to semantic information.Due to the incomparable feature extraction capability of the deep learning method on visual data,and the efficiency advantage of the hash-based approximate nearest neighbor retrieval method in storage and calculation,the combined deep hashing method has become a promising and effective solution for visual data retrievalBecause the output of the fully connected layer of the convolutional neural network has rich semantic information,most of the deep hashing methods directly use it as soft binary hash code and learn the hash function by encoding the global semantic information.But they lost the spatial information of the image,which may be one of the reasons for the performance bottleneck of model.When there are multiple instances in the image,we should pay attention to the local information which is highly relevant to the retrieval target,while the surrounding background is redundant and even interferential.Attention mechanism is publicly recognized as a good method to explore local information with high contribution to the task.It is a mechanism or methodology without strict mathematical definition.Many deep models use it as a branch network to create a space-wise or channel-wise mask,which is fused to the corresponding feature map of the main branch network and guided by minimizing the loss function of the main branch task.While the reason why it was related to the concerned object is exposited so subjectively.Given the above problems,this paper proposes a new explicit guided attention mechanism and applies it in a deep hashing method for image retrieval.On the problem of deep hashing method,our method avoids that simply using full connection layer features as approximate hash codes and ignoring the local information of the image.We use the attention mechanism to extract image features.As a result,the global semantic information and local information of the image are comprehensively taken into consideration.For the attention mechanism,we sublate the method which constructed the attention map in network branch and trained it with mainline tasks.It means that we first train it with classification tasks,which guided the spatial attention map having a clear semantic meaning.And then the attention map is used as auxiliary spatial information to help obtain local information on the pictureTo verify the effectiveness of the proposed model,we performed experiments on three datasets and compared the performance with state-of-the-art hashing methods The results of comparative experiments demonstrate that the retrieval performance of the method proposed in this paper is better than or comparable to several state-of-the-art methods.And the explicit guided attention mechanism proposed in this paper is an independent network branch that can be transplanted into other depth models in the CV field to improve the feature extraction capability of the model,which has high portability and practicality.