Research On Super-resolution Image Reconstruction Algorithm Based On Deep Learning

Super-Resolution reconstruction technology is a method to increase the pixel density of image from the software level without upgrading the hardware equipment.This technology can not only make images clear,but also is good at the following tasks.In recent years,deep learning is popular because of its powerful ability of learning and expressing complex features.The task of reconstructing high-resolution images from low-resolution images prediction is a ill-posed problem and requires complex feature extraction and transformation,while the deep learning is suitable for the task of image super-resolution reconstruction.This paper aims to improve the quality of super-resolution image by convolution neural network algorithm.The main work is as follows:1.The thesis analyzes and summarizes the research background of image super-resolution reconstruction and the research status at home and abroad,sorts out and summarizes the existing super segmentation algorithms and analyzes the common classical algorithms,and then introduces the related basic theory of depth learning and the evaluation method of SR algorithm in detail.2.The basic principle,advantages and disadvantages of SRCNN are analyzed.To solve the problems of small receptive field and slow network convergence,we propose a residual network structure using sub-pixel convolution layer.In order to prevent the gradient information from disappearing,residual connection is used.By using sub-pixel convolution layer in network,not only.artifacts can be avoided,but also unnecessary computation overhead can be reduced.Through experiments,we can see that the PSNR and SSIM indexes are improved.3.In order to make the SR image have more high-frequency information,the perception loss calculated by VGG19 is added to the generator's loss function,The experimental results show the proposed method can improve the detail clarity of the image and have more smooth and natural texture.