Research On Super-Resolution Reconstruction Algorithms For Infrared Images

Infrared imaging has the advantages of 24-hour imaging and intense penetration makes infrared images widely used in military surveillance,product testing,and medical examination,but due to the influence of external conditions such as infrared detectors make the existing infrared imaging systems generate infrared images generally have low resolution,easy to carry noise and blurred edge details and other deficiencies,so improve the resolution of infrared images gradually become an urgent problem in the field of image processing.There are two main ways to improve the resolution of infrared images:hardware and software.The high cost and limitations of the hardware method make the software method to improve the resolution of infrared images gradually become the primary choice.Therefore,low-resolution infrared images are studied in this thesis,and a software approach to improving the resolution of infrared images is investigated to address the existing problems of the SISR algorithm.The research for this thesis is as follows:(1)In response to the problem that existing super-resolution reconstruction algorithms for infrared images mainly use a large number of infrared images to train the network performance to improve the resolution of infrared images,the adaptive Gaussian filter is used to improve the interpolation-based super-resolution reconstruction algorithm in this thesis.The experimental results show that the algorithm can realize the effective improvement of infrared image resolution without external information,and the algorithm principle is simple and easy to operate.(2)In response to the existing visible image super-resolution reconstruction algorithm being more mature and the public infrared data set is not sufficient,the existing VISR algorithm is investigated through this paper to improve the VISR algorithm based on comparing the difference between visible and infrared images for deficiencies such as blurred texture details in infrared images.Based on the ESRGCNN algorithm with shallow structure combined with scale decomposition,the image is divided into a base layer and detail layer,and the histogram equalization and detail enhancement are carried out for each characteristic respectively,and then the final super-resolution infrared image is obtained by weight construction with the processed low-resolution infrared image.The experimental results show that the algorithm can effectively improve the objective evaluation index and the subjective visual texture detail clarity is improved.(3)For the practical application of infrared images,pairs are more difficult to obtain,pairwise mapping networks are selected for optimization in this thesis.To obtain more details of IR images,the attention mechanism is introduced into the pair mapping network and multi-scale detail filtering is used to improve the sharpness,which is combined with the swish function to enhance the super-resolution reconstruction performance of IR images in the dual mapping network.Experiments show that the optimized dual mapping network has improved super-resolution reconstruction performance under infrared dataset.Considering that the existing visible image and infrared image fusion algorithm will have the problem of unnatural shadow transition,the optimized dual mapping network is introduced into the multi-source image fusion algorithm.Through experimental verification,the performance of the image fusion algorithm with the introduction of the SISR algorithm has been significantly improved.