Research On Infrared Image Super-resolution Reconstruction Method Based On Residual Dense Network And Multi-branch Feature Fusion

Infrared imaging technology is not affected by external light,and has high sensitivity to temperature and longer detection distance,so it can be widely used in many fields.However,the resolution of infrared images is generally low,and in practical applications,high-resolution infrared images are more conducive to the completion of tasks.For example,high-resolution infrared images are conducive to the completion of target detection and recognition in security inspection,detection,anti-theft and night detection.Therefore,it is necessary to study the super-resolution of infrared images.In order to solve the problem of low infrared image resolution,this paper designs a SRGAN infrared image super-resolution reconstruction algorithm based on residual dense network.This algorithm uses features of multiple layered information to reconstruct super-resolution images.In order to enhance the capability of feature extraction,an infrared image super-resolution reconstruction algorithm based on multi-branch feature fusion is designed,which realizes multiple feature extraction.The main research contents and results of this paper are summarized as follows:(1)In view of the problems of SRGAN algorithm in the super resolution reconstruction of infrared images,such as the image is too smooth,lack of texture information,and the reconstructed image is not realistic enough,this paper designs an SRGAN algorithm based on residual dense network.Based on the SRGAN algorithm,the generate network and up-sampling in SRGAN are respectively improved.The hyperparameters of the loss function and the convolution kernel size are also discussed.In the generate network,the residual network of the generation network of SRGAN algorithm is changed into the residue dense network.While maintaining the transmission cascade features,the high frequency information required for reconstruction image can be better obtained by the fusion of the characteristic information of each layer by using the feature information of each layer.In order to improve the effect of super-resolution reconstruction,progressive up-sampling is adopted to extract features from the up-sampled images at each stage through convolution.Progressive up-sampling can reduce the difficulty of learning.The experimental results show that compared with the original high-resolution infrared image,the evaluation indexes of PSNR and SSIM were improved significantly..(2)The problems of EDSR algorithm in super resolution reconstruction of infrared images,such as artifacts,mesh phenomenon in the details of magnified images and unclear image reconstruction,etc.This method designs an algorithm based on multi-branch feature fusion by referring to the structure of EDSR network.The core of improvement is to change the feature extraction branch of EDSR from the original one branch to three branches.The first branch retains the residual network of EDSR,but the two 3×3 convolution kernels size is changed to 1×1 and 3×3 in the residual network,so as to reduce the number of parameters and save the training time on the premise of ensuring feature extraction.The second branch applies an improved Inception network with a convolution kernel size of 1×1,3×3,and 5×5,respectively,through different convolution sizes extract information of different scales.The third branch is strengthened the edge features.A general convolution module of convolution(1×1)-batch normalization-convolution(3×3)-batch normalization is designed.The edge images detected by the Canny operator are input into the network superposition of multiple ordinary convolution modules to enhance the edge features.This method adopts the L2 loss function.The L2 loss function has a greater penalty for the larger pixel difference value of the image,and a higher tolerance for the smaller pixel difference value,which makes the training more stable and the convergence state can be reached faster,and the reconstructed image is smoother.The non-reference evaluation indexes PI and NIQE is used to objectively evaluate the quality of reconstructed images.The experimental results show that the proposed algorithm can significantly improve the objective evaluation and subjective evaluation of reconstructed images with magnification factors of 2,4 and 8.