Research On Deep Hash Algorithm Based On Gradient Attention

With the rapid development of 5G technology,people are not only using the Internet for convenience,but also making the image data on the Internet present an explosive growth trend.It is an important task to find the image needed by users quickly and accurately in the direction of image retrieval.The Early text-based image retrieval technology is not fast,usually need to manually annotate the image,through the way of matching the text to retrieve the image.Hash based image retrieval technology can reduce data storage space and improve retrieval performance,so it is widely used in the field of image retrieval.With the further development of deep learning research,the performance of deep hash method in the field of image retrieval continues to improve.Most of the existing deep hashing methods are carried out under the condition of supervision,and the label information of the image is used in the process of network training.Most of the existing supervised depth hashing methods design the loss function according to the similarity between images,and constrain the hash code from the Euclidean space to the Hamming space.However,due to the lack of semantic information utilization,this constraint method may still produce sub-optimal hash code.Most of the massive images on the network are not labeled,so it is unrealistic to label these images manually,therefore,unsupervised hash method is worth studying.In view of the existing problems,the work of this paper is as follows:In order to solve the problem of sub-optimal hash code in supervised deep hash method,this paper uses multiple loss functions to constrain the generation of hash code and make use of more semantic information.The deep semantic hashing algorithm(DSHA)proposed in this paper is based on supervised learning.It uses pairwise tag information to learn hash codes in the end-to-end framework,and uses pairwise loss and contrast loss to regularize the real value output to approximate the required discrete value.In the unsupervised hash method,the existing unsupervised hash algorithm based on learning usually uses the pre-trained network to extract features,and then uses the extracted feature vector to construct the similarity matrix,and guides the generation of hash code by gradient descent.Previous studies have shown that the algorithm based on gradient descent will cause the hash codes of paired images to update to each other's positions in the training process.For unsupervised training,this situation will lead to a large fluctuation of hash code in the training process,which limits the learning efficiency of hash code.In order to solve this problem,this paper proposes a deep unsupervised hash algorithm based on gradient attention,which use pre-trained network models to extract image features,calculate the cosine distance of the corresponding features of the pair of images,and construct a similarity matrix through the cosine distance to guide the generation of hash codes,the gradient attention mechanism is added to control the change of the gradient in training process.