| High-resolution images have rich details and clear structure,which can meet the visual needs of human beings.However,limited by imaging device,people have difficulty to obtain high-resolution images.Super-resolution(SR)reconstruction is a technology that uses algorithms to recover high-resolution images from the corresponding low-resolution ones.According to the number of input images,it can be divided into single image superresolution(SISR)and video super-resolution(VSR).In recent years,with the development of convolutional neural networks and big data,super-resolution algorithms based on deep learning have gained widespread attention because their reconstruction performance far exceeds traditional ones.Inspired by this,this paper uses deep learning and information feedback theory to conduct exploratory and innovative learning and research from single-image to video-image.The main research work is as follows.Most of SISR algorithms based on deep learning tend to deepen the network,although it is helpful to improve the performance,it brings too much calculation.Moreover,the information flow is too long to preserve low-level information during training,which restricts feature learning ability.To solve these issues,a single image super-resolution algorithm based on multi-scale feedback residual network is proposed.First,the densely connected multi-scale residual blocks are designed to learn the nonlinear mapping between high and low resolution image spaces.Specifically,the multi-branch structure is used to extract context information of different scales,and the kernel selection mechanism is used to dynamically adjust the receptive field of output features.Then,the learned high-level features are used to improve the initial low-level features through the feedback structure.This feedback mechanism enhances feature representation ability without increasing the parameters.Finally,directly upsample high-level features to obtain high-resolution images.Experimental results show that the proposed algorithm surpasses most existing image super-resolution algorithms about the PSNR and SSIM scores on the five standard testing datasets while maintaining fewer parameters.Video super-resolution can not only consider the spatial information of the current reference frame,but also accurately and effectively utilize the temporal information of multiple neighboring frames.For this reason,combined with the effective multi-scale feature extraction and feedback mechanism in the previous chapter,a video super-resolution algorithm based on multi-stage feature fusion network is proposed.First,through the deformable alignment module,the motion of the reference frame and neighboring frames are adaptively compensated in the feature level before the super-resolution process,so as to achieve feature alignment between frames.Then,the aligned features of neighboring frames are used as a condition input,and are fused into the reference frame information in the form of modulation at different stages of the super-resolution network.Meanwhile,the feedback connection is used to enhance the spatiotemporal feature information,and finally the high-resolution image of the reference frame is directly reconstructed.Experimental results show that the proposed algorithm has a good reconstruction performance on the three video testing datasets.Compared with the most advanced approach,the PSNR values on the SPMC-11 and Vimeo-90K-T are improved by0.08 dB and 0.26 dB. |
