Research On Infrared Image Enhancement Algotithm Based On Deep Learning

With the development of technology,infrared images have a wide range of applications in various fields.These applications require a large number of high-resolution infrared images.Due to hardware limitations,current infrared images have low resolution and poor visual quality.Infrared image super-resolution reconstruction based on deep learning is simpler,more effective and universally applicable than traditional enhancement algorithms,and can solve this problem well.Therefore,in this paper,the method of super-resolution reconstruction is used to improve the quality of infrared images and obtain high-resolution infrared images.The main work of this paper is as follows:1.This paper introduces an iterative error reconstruction network based on SRDenseNet to improve the resolution of infrared images.This paper firstly implements efficient dense connections based on linearly compressed skip connections.Efficient dense connections sacrifice performance to a certain extent,but can significantly reduce the parameters and computation of ordinary dense connections.On this basis,an iterative error reconstruction module is introduced to improve the performance of the network,enabling the network to recover more textures and edges.The specific iterative process is as follows:First,obtain the initial highresolution image,advanced features and up-sampling features.Second,a high-resolution error image is obtained by reconstructing the error between the initial high-level features and the upsampled back-projected features.Finally,the error image is added to the initial image to get a new highresolution image.The above process is iterated,and when the number of iterations reaches the iteration threshold,a high-resolution image will be obtained.2.Although an iterative error reconstruction network is introduced to restore texture detail information,there are not many high-frequency information components in infrared images,and the network only uses single-modal input of infrared images,which leads to the lack of highfrequency detail information in the network and the visual effect of the reconstructed image.not good.In order to further improve the performance of the network and the visual effects of reconstructed images.Based on the iterative error reconstruction network,this paper proposes an infrared image superresolution network with multi-modal input of IR images and corresponding RGB images.The network first denoises the input low-resolution infrared image through a guided filter;then a feature extraction module is designed to extract features from RGB images,and sub-pixel convolution is used instead of deconvolution to improve the accuracy of feature extraction.;Finally,combine the RGB feature image with the IR feature image to reconstruct the infrared image.The experimental results show that the final network in this paper has obvious advantages compared with the iterative error reconstruction network and the existing network.