Research On Hashing Algorithm Based On Deep Learning

In recent years,the rapid development of the Internet has led to an explosive growth in image data for each industry and every firm.It has become a hot topic to retrieve the images needed by users efficiently and accurately in a large number of image data,which has been widely concerned by researchers.Faced with an urgent need for image retrieval technology in the industry,text-based image retrieval was born.Considering that the visual characteristics of images are also important for image retrieval,content-based image retrieval techniques was proposed.Text-based image retrieval technology needs a lot of manual annotation,which is time-consuming and difficult to achieve for massive data retrieval;However,content-based image retrieval technology has the following two problems: the problem of semantic gap and the problem of dimensional disaster.This paper combines deep learning and hashing algorithms to solve the problems bring about by the above image retrieval technology.Deep learning can automatically perform feature learning and has strong capabilities,while the hashing algorithm can map the high-dimensional content features of the image to the low-dimensional Hamming(binary)space,which effectively reduces the storage space of the image in the computer.This paper mainly uses the spatial attention mechanism and N-pair loss to improve the deep hashing algorithm respectively,and performs correlation verification through two benchmark datasets.(1)For the problem that most images will be affected by geometric distortion and background factors,and in the past,deep hashing methods only used global images to generate binary hash codes,and did not consider local spatial attention information,which is also important for image positioning and object detection.Therefore,a novel deep hashing framework is proposed in this paper,which integrates the local spatial attention mechanism and global image information mechanism into an end-to-end network structure.To make the hashing algorithm more efficient,the proposed framework includes two subnets.A sub-network focuses on the specified location of the target object,and it uses a local spatialattention mechanism.Another sub-network uses a global convolutional neural network to extract the global features of the image.Finally,The image features output by these two subnets are fused to form the final binary hash codes.(2)Aiming at the problems of the existing deep hashing algorithm based on contrast loss and triplet loss: the network is simple;it uses only one negative example,does not interact with other negative classes in each update,and convergence is slow.In this paper,we use a novel hashing algorithm called N-pair loss-based deep residual hashing algorithm to solve this problem.To reduce the computational burden,we use batch construction.The proposed objective function allows a joint comparison among multiple negative examples to summarize the triplet loss,and can interact with other classes of samples at each update to achieve the global optimal solution.This paper conducts experiments on two benchmark image datasets(CIFAR-10 dataset and NUS-WIDE dataset).The experimental results show that the improved deep learning-based hashing algorithm in this paper can better meet the requirements of mass image retrieval,and verify the effectiveness of the improved scheme.