Research On Infrared Image Super-resolution Reconstruction Method Based On Transformer Network And Feature Fusion

Infrared imaging is a technology that converts infrared radiation energy,which is not visible to the naked eye,into a visible image for people after being captured by an infrared detector.This technology has the characteristics of good concealment,strong penetration,and high recognition,therefore,it is widely used in military,security,and surveillance fields.However,the resolution of the obtained infrared images is generally low due to the limitation of infrared wavelength and shooting equipment,which seriously limits the application and development of infrared images in various fields.A hardware approach to obtain high-resolution infrared images requires complex manufacturing processes and high costs;while a software approach,through image super-resolution reconstruction techniques,can significantly improve the quality and clarity of infrared images and is easier to implement.Therefore,in this paper,two single-frame infrared image super-resolution reconstruction algorithm models are proposed for the characteristics of infrared images and the problems of existing deep learning-based image super-resolution reconstruction methods,respectively.(1)To address the problems of blurring,lack of details,and unclear edges in infrared images,this paper designs an edge-enhanced cross-shaped window Transformer super-resolution model for Infrared Image Reconstruction based on the Swin IR.The Transformer layer of the cross-shaped window in ECSSR can calculate the horizontal and vertical streak self-attention in parallel,and obtain a better reconstruction effect without increasing the amount of calculation.In addition,an EDAN edge detection auxiliary network is proposed by combining Bicubic and RCF to obtain more accurate edge information,which can be used as supplementary information to enhance the reconstructed infrared image edges.The experimental results show that the proposed method outperforms the current representative methods in terms of quantitative indexes and visual effects,and the edges and textures of some reconstructed images are still clearly visible even when the reconstruction magnification is ×8.(2)To address the problems of insufficient use of RCAN features in the superresolution reconstruction model,this paper proposes an infrared image super-resolution reconstruction model FFCASR based on feature fusion and coordinate attention.The multi-scale shallow feature extraction module is introduced in this model,which can obtain more abundant shallow features from infrared images.Meanwhile,a global feature fusion module is proposed to address the problem of underutilization of the output features of the residual group in RCAN.This module can fully reuse the output features of each residual group,dig deeper into mapping relationships,and prevent information loss.In addition,by using coordinate attention,the model can capture remote dependencies in one spatial direction;and preserve accurate location information in another spatial direction.The results of comparison experiments and ablation experiments show that FFCASR can effectively improve the ability of learning and expressiveness of the model,and the reconstructed infrared images have more detailed information.