Research On Image Retrieval Based On Hash Algorithm And Generative Adversarial Network

In the 21 st century,with the continuous development of the current hardware equipment,big data technology and so on,in addition to the existence of some proprietary image databases on the network,people through information exchange,e-commerce and so on make the Internet produced a variety of image data,and every day by hundreds of thousands or even millions of growth.Image retrieval in computer vision is faced with the problem of how to label the large-scale image data and how to find the pictures that users need conveniently and quickly.In content-based image retrieval,the traditional method is usually to manually annotate the image data,and the process of manual annotation will be affected by the individual of the annotator,resulting in the difference of images,and the workload is huge.In addition,the retrieval is usually carried out one by one matching retrieval according to the features extracted from the images.The shortcoming of this algorithm is that the extracted image feature dimensions are too high.In a large-scale image database,the storage of these high-dimensional features will consume too much space and the retrieval speed is too slow.However,the approximate nearest neighbor hash algorithm can project the data in the high-dimensional space into the Hamming space as binary code through the hash function transformation,so as to reduce the data dimension and reduce the storage space.It can also compare the image similarity through the Hamming distance in order to improve the efficiency of image retrieval.In addition,generative adversarial network in deep learning can generate pseudo-images similar to real images after training,which can be considered to have the same label and can be used to guide unsupervised learning.However,at present,hash-based image retrieval requires several times of training to generate hash codes.Image retrieval based on generative adv network still has a large research space in solving the semantic gap and understanding the distribution of visual features of images.Aiming at these problems,the main work of this paper is as follows:In order to solve the problems of unstable training using original generative adversarial model and inadequate semantic learning of small size single label images,attention mechanism and generative confrontation network were introduced into deep hash learning,and an image retrieval framework based on attention mechanism and improved GAN was designed.Binary attention generation versus image retrieval model.First,an encoder network is constructed to learn the most critical features of the image by adding the attention mechanism CBAM.Then,the pseudo-label matrix is constructed,and the hash code is generated through the hash layer and input to the generator to reconstruct the original image.Finally,the true and false samples are discriminated through the discriminator.In the process of adversarial learning,the quality of generated hash code is improved,so as to improve the performance of image retrieval.In order to verify the effectiveness of the proposed model in the field of image retrieval,this paper uses two commonly used evaluation indexes to carry out relative comparative experiments on public data sets,which shows that the model does have outstanding performance in terms of retrieval accuracy and retrieval effect.At the same time,through the ablation experiment,it is found that the innovation points proposed in this paper can solve the above-mentioned problems to a certain extent.Therefore,the image retrieval algorithm and related solutions proposed in this study provide reference for unsupervised large-scale image retrieval and have certain practical application value.