Study On Transforming Ir Images To Daytime Natural Color Images Using Adversarial Generative Netwrok

Infrared images are used in low-light conditions and night vision environments,and play a major role in driving system,security monitoring,and military exploration.It reflects the radiation information of the target scene.Compared with the visible light image,the information is seriously missing,especially the color information,and improving the quality of the collected image requires a huge cost.According to the biological vision mechanism,humans can distinguish more than twenty different gray scales from black to white,and the cones can distinguish thousands of different degrees of color,so the human eye is far less sensitive to grayscale images than color images.In addition,the infrared image is darker than the grayscale image,and the image cannot directly convey the real situation.The prediction of daytime natural light scene by nighttime infrared image is a very typical unsupervised image translation problem in deep learning.This paper mainly uses the unsupervised generative network model to solve the problem of generating nighttime infrared image as daytime color image.This paper first analyzes the characteristics of traditional infrared night vision colorization,and then finds that there are few methods involved in the daytime colorization of night vision infrared,and the generation of network to solve similar problems better.However,aligned datasets of nighttime infrared images and daytime color images are scarce,so we choose to use unsupervised learning to train unpaired night vision infrared and daytime color image sets.The third chapter of this paper mainly introduces the cycle-loss based generative network,and uses the cycle-loss to train generative network to transform the infrared to daytime color image and display the results.The fourth chapter uses the generative network with the variational auto-encoding structure and the cycle loss to generate the color images.The network is based on the assumption that there is hidden space in the margin distribution in the conversion process of two domains.At the same time,the gradient information and luminance information of the image are extracted,and the parallel structure of discriminators is introduced.We set separately the gradient,brightness,and Gaussian blurred of one same image as input to each discriminator.In the training process,the correlation parameters are used to optimize the network parameters.Based on the original loss,the Laplacian operator is introduced to extract the approximate boundary of the color image and the original infrared image,and the boundary loss is calculated.In the colorization task,the method of using the enhanced gray cycle loss and appropriately weakening the cycle loss of two colored images is used to avoid the image generating colors being excessively single.The results show that the background effect of generating scenes based on auto-encoding unsupervised generative model is better than the generation foreground,and the added structural loss is helpful for generating background.Using SSIM to illustrate the introduction of Laplacian-based structural loss is better than using only correlation loss.Finally,several network modification schemes are proposed,such as adding parallel decoding network in the generation network,simultaneously inputting multi-scale images and calculating loss during training to optimize the resolution of generated images,introducing global attention mechanism to optimize generative network.The spectral normalization method can be used on it.