Research On Dissimilar-source Image Transfer And Enhancement Based On Generative Adversarial Network

In recent years,the multi-source image sensing technology has developed rapidly.These sensors can sense the visible spectrum,infrared spectrum,and even the depth information of space,which greatly expands the perception capabilities of computers.Infrared cameras can work in low-lighting conditions and harsh weather conditions,making them widely used in medical imaging,video surveillance,military guidance and other fields.Infrared sensors can sense the temperature of an object through the amount of infrared radiation energy emitted by the object,and also play an active role in the prevention and control of the epidemic.However,the working principle of infrared sensors leads to some inherent deficiencies in infrared images.First of all,there is a large modal difference between infrared images and visible light images,and the human eye is far less sensitive to gray values than color images.Furthermore,due to the limitation of infrared sensor performance,infrared images have shortcomings such as blurred edges,large noise,low resolution,and lack of details.Therefore,in response to the above two issues,this article takes infrared images as an example to carry out research on multi-source image transfer and enhancement technology.1.Research on dissimilar-source Image transfer technology.Aiming at the shortcomings of the existing infrared image transfer methods,this article designs an infrared image transfer algorithm STVGAN based on semi-supervised generative adversarial network.STVGAN uses a semi-supervised training method to design an efficient generation network and a discriminant network and a compound loss function.Based on a large amount of unsupervised training data,only a small number of supervised learning samples can achieve the transfer result of the existing supervised learning method.Compared with the infrared-visible image transfer algorithm Tir2 Lab,STVGAN has improved the PSNR and SSIM indicators by 2.71 and 0.1,respectively,which is also significantly improved compared with the training results of the existing image transfer network.2.Research on dissimilar-source image enhancement technology.Aiming at the problem of the domain gap between the low-resolution images used in the training of the infrared image super-resolution algorithm based on supervised learning and the real low-resolution images,this article studies an infrared image enhancement algorithm DSRGAN based on unsupervised degradation learning.The article divides the image enhancement network into two modules: unsupervised degradation learning and image super-resolution reconstruction.The degradation network module is designed based on the cyclic generation adversarial network.The image super-resolution reconstruction network improves the ESRGAN super-resolution reconstruction algorithm.Through subjective and objective comparison experiments,it is proved that the unsupervised learning algorithm designed in this article has better performance in super-resolution reconstruction of real low-resolution infrared images than the super-resolved learning algorithm that obtains low-resolution images through down-sampling,and its performance is close to directly obtaining real low-resolution images.The supervised learning algorithm for high-resolution images does not rely on pairs of high-and low-resolution images.3.Network training optimization.Nowadays some high-performance networks represented by generative adversarial networks have increasing demands on GPU computing power,video memory,and bandwidth,and the lack of computing resources has also become an obstacle to the research work of this article.This article uses the advantages of low-precision format in memory layout and computational efficiency,uses automatic mixed-precision technology to accelerate training and experimentally proves the advantages of automatic mixed-precision training.