Research On Single Image Super-resolution Reconstruction Algorithm Fusing Frequency Domain Information

In the era of big data,digital images have become more and more widely used in various fields.For example,in special fields such as security and medical care,high-quality images play a crucial role in the judgment of professionals,while in social media,e-commerce and other fields,high-quality images can provide users with better application experience.However,in the practical application of the real world,the hardware equipment is affected by the cost,natural environment and other factors,some images with low resolution,poor quality,and blurred texture details are collected,improving the quality of this part of the image is an urgent problem to solve.Therefore,it is of great research significance and application value for various fields to use image super-resolution technology to carry out high-quality image reconstruction to improve the indispensable rich details and reliability.In this thesis,the key technologies of image super-resolution are studied based on deep learning,focusing on the reconstruction of high-resolution image texture details.The main research contents of this thesis are as follows:(1)A dual-stream network super-resolution algorithm that combines spatial and frequency domain information is proposed to improve the accuracy of image reconstruction.As the super-resolution algorithm based on deep learning mainly learns the nonlinear mapping relationship in the spatial domain of the image,the high-frequency information such as edge and texture occupies a relatively small proportion in the image,which leads to less high-frequency information entering the network and weak learning and reasoning ability,leading to blurred details of the super-resolution reconstructed image.Therefore,this thesis introduces image frequency domain information to enhance the learning and reasoning ability of image high frequency information in the network.The algorithm is composed of spatial and frequency domain branches.The spatial branch network mainly learns low-frequency information such as image profile and contour,while the frequency domain branch network mainly learns high-frequency information such as image edge and texture.In addition,a frequency domain loss is proposed to guide the frequency domain branch to focus more on the learning of high-frequency details.The experimental results prove that the edge texture of the reconstructed image is clearer in subjective evaluation,and higher image reconstruction accuracy is obtained in objective evaluation,which verifies the effectiveness of the algorithm proposed in this thesis.(2)A lightweight super-resolution algorithm with adaptive channel pruning is proposed to improve the speed of image reconstruction.The algorithm uses a weight gate mechanism to learn the relative importance of frequency domain channels in an adaptive manner,and remove channels with low contributions without affecting the generalization ability.The algorithm effectively reduces the model scale and computational complexity,increases the operating speed,and eliminates the interference of redundant frequency domain channel information,making the network more efficient and focusing on learning image details.The experimental results prove that the algorithm not only effectively reduces the network scale,but also obtains better subjective and objective evaluation results.(3)The high-quality image automatic detection super resolution system is designed and developed to improve the efficiency of image reconstruction.The system uses the NIQE^[56]algorithm for unsupervised image quality evaluation,and automatically selects low-quality images based on the evaluation results.At the same time,it increases the manual filtering function to improve the flexibility and reliability of the system's image quality detection.Finally,the lightweight super resolution algorithm of adaptive channel pruning proposed in this thesis is used to reconstruct low-quality images to obtain high-quality images with clear texture details.The experimental results prove that the system meets engineering requirements design,improves efficiency,and reduces labor costs.