Deep-learning Based Image Super Resolution Algorithm In Frequency Domain

With the development of smart devices,people have more and more demands for clear and high-resolution images.However,in some applications such as security monitoring and medical images,due to factors such as shooting environment,imaging equipment and storage process,ideal high-resolution images are often difficult to obtain.This leads to that the images appear before our eyes are mostly low-quality images that are polluted or have lost pixels.In order to improve the image quality without changing the existing hardware equipment and external environment,the image super-resolution technology was proposed.Image super-resolution technology refers to reconstructing high-quality HR images from low-quality LR images through algorithms.Image super-resolution technology is not only low-cost,but also has low requirements for equipment and environment.It has been widely used in image communication,security monitoring,medical diagnosis,remote sensing imaging and other fields.Currently,learning-based image super-resolution algorithms have been widely studied due to their fast reconstruction speed and high reconstructed image quality.However,many learning-based super-resolution algorithms do not consider the inherent frequency features of images and cannot effectively recover highfrequency details,resulting in poor visual effects of reconstructed images.The purpose of this paper is to reconstruct more texture and detail information on the premise of ensuring that the reconstructed image is not distorted,improving the quality of the reconstructed image.The main innovative studies are summarized as follows:(1)Aiming at the problems of cumbersome feature extraction,high computational complexity,and inability to effectively distinguish different image block textures in the mapping-learning based algorithm,this paper proposes an image super-resolution algorithm based on frequency-decomposition mappinglearning.This algorithm makes full use of different frequency information of image blocks,and uses decision-tree structure to classify image blocks in the order of low frequency to high frequency.This classification mode can better distinguish image block textures of different frequencies and different directions,improve the feature similarity of each class,and help mapping kernels to better fit the image block characteristics of each class.Therefore,it can more accurately reconstruct high-frequency details that are lost in LR images.This algorithm is simple,robust,and has great practicability.(2)Deep learning-based super-resolution algorithms only focus on endto-end mapping learning,but ignore related domain priors,resulting in that the high-frequency information of reconstructed images cannot be effectively recovered.In order to solve the above problems,this paper proposes a frequencyattention-based image super-resolution algorithm.This algorithm explicitly presents the spatial features of different image frequencies to CNN,and enhances the limited high-frequency information of LR space through the attention mechanism,giving full play to the complementary and guiding role between the high-frequency and low-frequency information.This algorithm makes the network to put more resources on the important high-frequency information of LR space,which can reconstruct more high-frequency details and improve visual effects.The algorithm can recover more visually pleasing highfrequency details with fewer network parameters and faster runtime.