Image Super-resolution Restoration Based On Deep Learning

Image super-resolution reconstruction has been a hot issue in the field of image processing.Now as the preferred method for image super-resolution reconstruction,deep learning-based methods have made great research progress in the fields of video surveillance,remote sensing satellites,medical imaging and multimedia.Therefore,this thesis focuses on the research of image super-resolution reconstruction method based on deep learning.By effectively designing the structure of convolutional neural network,the effect of image super-resolution reconstruction based on deep learning has been significantly improved.However,this method does not make full use of the non-local correlation in the natural image and does not make effective use of the higher-order statistical features of the image,resulting in an unsatisfactory reconstruction performance.To address the above problem,this thesis proposes a deep network based on non-local and second-order feature fusion for image super-resolution.On the one hand,non-local module fully uses the self-similarity characteristics as the prior information to exploit the rich structure information in the natural image,so as to obtain the spatial context information.On the other hand,the second-order attention mechanism can implement the adaptive weighted processing of the characteristic response intensity between channels,and explores the second-order characteristic statistics through the covariance pooling operation,so as to achieve stronger feature expression and correlation learning ability.In addition,in order to effectively utilize the abundant low-frequency information,long skip connection is applied in the network to pass more abundant low-frequency information from LR images and ease the network training.Experimental results verify the influence on the reconstruction performance of non-local and second-order feature fusion module,and demonstrate that the proposed method can achieve more desirable performance over several state-of-the-art methods in terms of quantitative metrics and visual quality.By capturing the inherent attributes of natural images,the feature expression ability of the network is effectively improved.However,this method only extracts image features in single scale level.In fact,the image feature information is usually in different scales,which leads to the loss of complete feature information.To solve the above problem,this thesis introduces a multi-scale feature fusion module,and proposes a deep network based on attention mechanism and multi-scale feature fusion for image super-resolution.As for attention mechanism,the non-local and second-order feature extraction modules can be regarded as spatial attention mechanism and channel attention mechanism respectively.And as for multi-scale feature fusion module,it uses the diffetent scale convolution kernel to extract the information of different scale patches,so as to preserve the complete multi-scale features.Experimental results show that the proposed method is more efficient and reliable over other representative algorithms in image super-resolution reconstruction.