Reasearch On Infrared Image Generation Algorithm Based On GAN And Physical Model

Infrared thermal imaging is imaging by reflecting the temperature field on the surface of the object,so it has the advantages of prominent heat source objects,not affected by lighting conditions,and certain antiinterference ability to fog and haze.Therefore,Infrared images can provide useful information in many situations where visible images are difficult to use,such as in low-light conditions at night,where there is smoke.However,the acquisition of infrared image requires professional and expensive equipment,and has high environmental requirements,so the existing infrared image dataset is not rich,which greatly limits the research and application of computer vision technology in infrared image.The purpose of this study is to research the infrared image generation algorithm to generate infrared images from visible images.In this thesis,a supervised image-to-image translation method is used to generate infrared image data from existing abundant visible image data.First of all,this thesis proposes a new dataset LLVIP to address the scarcity of existing visible-infrared paired datasets.The dataset is a visibleinfrared paired image dataset under low light conditions,containing 30976 images,namely 15,488 pairs of images captured in low light conditions,all of which have been registered and pedestrian annotated.The dataset is compared with other visible-infrared datasets,and the performance of image fusion and pedestrian detection algorithms on this dataset is evaluated,which proves the high quality and importance of the proposed dataset.This thesis proposes a new infrared image generation algorithm based on Pix2pixGAN and physical model.Based on the prior knowledge of physical model of infrared thermal imaging,this thesis proposes weighted loss of heat source mask,which improves the quality of key areas of generated images.Based on the consistency of feature space,the consistency loss of shallow feature is proposed to constrain the edge and texture of the generated image.This thesis also improves the generator structure by adding attention mechanism to improve the generator’s ability to discover key areas and further improve the quality of the generated results.Experimental results show that the infrared image generated by the above improvements is closer to the real image.