Learning To Hash Based On Generative Adversarial Networks

The rapid development of the Internet has brought us a huge amount of data,and people's demand for data is also increasing.Among them,efficient retrieval for large-scale image data sets has become an indispensable requirement for contemporary social life,and it is also an important research direction in the field of computer vision.At the same time,with the rise of deep learning technology,the idea of learning a complex hash map from a large amount of data to reduce dimensionality and retrieve data is well realized.The difference from the traditional method is that it can fully learn the visual and semantic features of the image from the data and serve the hash learning.The coding effect is obviously superior to the artificially formulated features and hash functions,which can be quickly obtained short and Accurate binary encoding.However,hash learning requires a large amount of data to train the model,and the deep learning supervised hash algorithm relies on a large amount of training data and its annotation.If the training data is insufficient,the model is easy to over-fit and cannot be put into practical use.In reality,the cost of labeling a large amount of data is extremely high.In addition,there are problems such as difficulty in training deep hash models and difficulty in avoiding quantization errors.Therefore,this paper proposes two image hash learning models based on generating confrontation networks,trying to use GAN's powerful generating ability to assist the hash learning process,to solve the above problems to some extent.This paper first proposes a two-stage model,referred to as ACGANH.This model is built based on the classic supervised generation against the network ACGAN.The first stage pre-trains an ACGAN network,and the second stage uses the pre-trained D network parameters to initialize the parameters of the hash network for the migration learning of the hash task.The G network is used as a data enhancement tool to complete network training.ACGANH can automatically obtain tagged data to assist hash network training and make improvements in hash network initialization.Experiments show that our method has greatly surpassed the baseline method to achieve better performance.In addition,we have also carried out semi-supervised learning improvements and achieved certain results.This paper also proposes an end-to-end hashing model that uses generation and confrontational design,referred to as GGANH.This model consists of three parts: encoder E,generator G,and discriminator D.It combines a well-designed loss function and uses a progressive scheme for training.In the process of progressive training,we can gradually train a high-performance image hash model to solve the difficult training problem caused by the complexity of the model.Experiments show that GGANH has done a good job of feature extraction,hash coding,image reconstruction and simultaneous learning against four tasks.The experimental results prove the global optimization ability of end-to-end training and improve the final retrieval performance.