Research On Infrared Super Resolution Algorithm Based On Generative Adversarial Network

Although the development of infrared imaging technology is very rapid,the application fields are very wide,and the resolution of imaging images is still very low.At this stage,the resolution of infrared focal plane arrays is mostly below 640x512,which not only fails to meet the needs of human eyes for high-resolution images with abundant details,but also cannot meet the display needs of imaging displays with 1K and 2K resolutions.Therefore,it is urgent to take measures to improve the resolution of infrared image.Image interpolation is the simplest measure to improve the resolution.When the image is magnified to the resolution of the imaging display,the details of the resulting image are blurred because no new information is introduced.Improving the resolution by increasing the array density and array size of the detector not only increases the cost,but is not suitable for large-scale amplification.Therefore,this paper focuses on the superresolution reconstruction technology and researches the super-resolution reconstruction technology of infrared images based on the generative adversarial network.The specific content is as follows:(1)An improved infrared super-resolution algorithm based on ESRGAN.In order to solve the problems existing in the current ESRGAN algorithm in the infrared superresolution reconstruction task,such as the lack of high-frequency detail of the resulting image,the generation of "artifacts" and the problem that the objective evaluation index does not reflect the quality of the image reconstruction well,the RCAGAN algorithm model is proposed.The image reconstructed by this model not only has better visual quality than commonly used super-resolution models including ESRGAN,but also the objective evaluation index PSNR/SSIM value reached 30.86 d B/0.8781,28.97 d B/0.8104,29.76 d B/0.8385 on the public data sets FLIR,CVC-14,and outdoor collection data sets respectively,which is higher than the commonly used super-resolution model at present,and higher than the ESRGAN algorithm index 3.02 d B/0.1092,2.68 d B/0.1199 and2.58 d B/0.0934;It is concluded that the RCAGAN super-resolution algorithm is better than the commonly used super-resolution algorithm.(2)Infrared image super-resolution algorithm of generative adversarial network based on blur kernel generation.Aiming at the problem of poor reconstructed image quality caused by the difference between the "ideal" downscaling kernel and the actual downscaling kernel,this paper proposes a new data set production model Kernel SR,and constructs a new infrared super-resolution sample library on the basis of this model.Commonly used super-resolution models have good performance on this sample library and can be generalized to actual scenes.The image reconstructed by this model not only has better visual quality than the commonly used super-resolution model(Kernel GAN+different algorithms),but the objective evaluation index PSNR/SSIM value reached 30.58 d B/0.8847,29.94 d B/0.8167 and 30.17 d B/0.8627 on the public data sets FLIR,CVC-14 and infrared super-resolution sample library,which is higher than the best super-resolution model based on blur kernel(Kernel GAN+ESRGAN)at present,4.23 d B/0.2036,3.12 d B/0.3029,5.44 d B/0.3076.It is concluded that the proposed superresolution model is a better model.Although the current version of the infrared image sample library is constructed using only three cameras,the trained super-resolution model shows good generalization capabilities for infrared images captured by other types of camera devices.