Research On Infrared Image Super-Resolution Algorithm Based On Deep Learning

Infrared imaging has many advantages,such as good anti-interference capability,long detection distance,and strong penetration ability.It provides valuable information for all-weather intelligent monitoring,medical imaging,military target detection,remote sensing,and so on.However,due to non-ideal optics and limited detector size,the infrared image resolution provided by the imaging system is low,making it difficult to meet the visual requirements of the human eyes.This limitation hinders the further development of infrared imaging,making it crucial to enhance the resolution and quality of infrared images.Currently,there are many super-resolution algorithms based on deep learning,but majority of them are designed for visible images,with less attention given to infrared images.Compared to visible images,infrared images have lower contrast,more apparent noise,and less high-frequency information.Therefore,this thesis focuses on the characteristics of infrared images and is dedicated to studying super-resolution algorithms for infrared images based on deep learning.The research work in this thesis is as follows:(1)Because of the characteristics of blurred edges and high noise in infrared images,it is difficult to fully obtain effective feature information from low-resolution infrared images.In order to solve the problem of blurred edges in infrared images,this thesis proposes a multi-scale edge attention residual module combining the characteristics of Res Net and attention mechanism.The multi-scale convolution kernel is used to ensure that the network can fully learn a variety of feature information.Meanwhile,the edge attention mechanism is used to increase the channel weight of the edge information,so as to better retain the edge information in infrared images.In order to solve the problem of large noise in infrared images,this thesis adds the total variation loss as a regularization term to the mean absolute error loss when designing the network loss function,so that the reconstructed image has spatial smoothness and suppresses effects of noise in infrared images.Finally,an ablation analysis of the proposed module and loss function is carried out,and the experimental results show the effectiveness of the module and the gain brought by adding the total variational loss.(2)This thesis improves the above network by introducing the high-frequency details of visible images through the correlation between infrared images and visible images,so as to effectively solve the problem of lack of high-frequency detail information in infrared images.Firstly,the structural similarity index and correlation coefficient are used to calculate the correlation between high-frequency information in the training dataset of visible and infrared images,resulting in correlation coefficient values of 0.713 and 0.792,respectively.This demonstrates the feasibility of using visible images to assist in infrared image super-resolution reconstruction.Then,based on the previously proposed network structure,a branch for extracting high-frequency information from visible images is added.The extracted high-frequency features from visible images are fused with the high-frequency features from infrared images to enhance the high-frequency details of infrared images.Finally,through objective metrics and subjective visual effects,experiments show that the improved algorithm in this thesis performs well at different sampling factors(x2,x3,and x4).