Infrared Image Simulation And Infrared Image Super-resolution Reconstruction Based On Deep Learning

Infrared imaging technology has many research values in many field,so many countries invest a lot in acquiring infrared data.The way of generating infrared scene by computer simulation has many advantages such as saving manpower,material resources,and financial resources.However,the traditional way of infrared-scene simulation by computer,needs a lot of physical calculation,and it is difficult to generate large-scale infrared scene by this way.In recent years,many researchers introduce machine learning and deep learning methods into the field of infrared scene generation.This paper uses the deep learning methods to convert the visible light images into the infrared images.And this infrared images will provide data support for the infrared images detection and recognition tasks.In order to carry out infrared target detection and recognition tasks under higher resolution imaging conditions,this paper decides to use the deep learning methods to super-resolution reconstruction of the low-resolution infrared images to get the infrared images with richer information and higher resolution.This paper uses the CycleGAN model as the basic network structure to generate infrared images.For the case that it is easy to occur over-fitting in the process of generating infrared images,which will lead to problems such as poor overall image quality and local error mapping in some generated infrared images.This paper proposes to add suppressor network structure between the generator network and the discriminator network.For the case that the result generated by one path in CycleGAN network may be different from the result reconstructed by another path.This paper proposes to add a loss function to calculate the path loss between the two training paths.By this way,the network can compare the result of the generator with the result of another path loop reconstruction.For the case that the original CycleGAN focus on the fine-grained features of image is not enough during the convolution operation,this paper propose to introduce the attention module(CBAM)into the network,to enhance the description of details in the generating process.In practical engineering applications,the infrared images obtained by traditional infrared imagers are low resolution and poor quality.Using these data as the training samples,may generate the infrared images with low resolution and low image quality.In order to perform infrared target detection and recognition tasks under higher resolution imaging conditions,this paper use the SRGAN model as the basic network to perform super-resolution reconstruction of the infrared scene.In order to retain the underlying information of the image,and prevent gradient dispersion and gradient explosion of the network,this paper use dense links to increase the entire network's perception of the underlying information.For the case that the edge and texture information is very important in infrared images,this proposes to add a loss function to calculate the edge gradient loss term into the loss function of the SRGAN network.By this way,the generator can fully learn the edge information of the infrared image,and gets the infrared images with clear and prominent geometric edges and reconstructed textures.