Research On Image Super-Resolution Algorithm Based On Traditional Feature And Convolutional Neural Networks

Image Super-Resolution(SR)aims to restore clear high-resolution images from low-resolution images.In engineering applications,due to limitations in capture conditions,usually only low-resolution images can be acquired.However,with the rapid development of electronic information technology,the resolution of terminals such as smartphones and personal computers is becoming higher and higher.Therefore,it is necessary to perform resolution enhancement on the received low-resolution images to adapt to the resolution of the terminal,making SR one of the hot research topics in the field of image processing in recent years.Currently,deep learning-based image SR is the mainstream of academic and industrial research.This type of method directly learns the mapping from low-resolution to high-resolution images end-to-end based on a large number of samples(low-resolution and high-resolution image pairs).Although deep learning methods can achieve effects that traditional methods cannot match,there are problems such as a large demand for training samples,high computational costs,and poor interpretability.Combining traditional methods with deep learning methods can partially alleviate these problems.Currently,most SR network structures are relatively deep.From the perspective of feature extraction,deep networks describe high-level semantic features more and provide less information on low-level features.This leads to the problem of distortion of low-level features in most SR networks,especially in regions with rich low-level features such as edges and textures.In order to provide SR networks with richer image underlying features,we study the combination of traditional low-level features and Convolutional Neural Networks(CNN)in SR algorithms,as follows:(1)Image SR network based on gradient-aware.Image information originates from changes in brightness and color in the spatial domain,and gradient is a fundamental tool for measuring changes.This makes gradient one of the most commonly used features.Some researchers have attempted to concatenate gradient domain features with CNN features in SR networks.This approach can bring some gains,but domain conflicts can affect the quality of the reconstructed images.To alleviate this problem,this thesis adopts an effective feature fusion method to avoid domain conflicts.The designed network includes a SR branch and a gradient branch,and the two branches exchange information through a gradient fusion module.At the same time,by adjusting the kernel sizes on the two branches,the feature maps maintain consistent receptive fields at the same depth.Numerical experiments show that the proposed SR network can achieve comparable restoration quality with the selected algorithms under the conditions of smaller parameter size and computational cost.(2)Image SR network based on gradient and edge awareness.In order to make the residual learning of the network pay more attention to the restoration of high-frequency information of the image,edge features are introduced on the basis of the gradient-aware SR network.A separate edge branch is designed in the network to extract edge-depth features,which are applied to the SR branch and combined with the image-domain feature maps to generate a spatial attention mask.Benefiting from spatial attention masking can generate more discriminating feature maps at the edges of the image.In addition,pseudo-artifact detection loss and edge loss are introduced.Perception-driven SR algorithms often generate some artifact at sharp edges of the image.The artifact detection loss determines whether the current pixel produces artifacts based on variance,and adds pixel loss at pixels where artifacts may occur to mitigate the generation of artifacts in the SR network.The purpose of edge loss is to enhance the quality of recovery in the center of the edges.Numerical experiments show that introducing edge features,pseudo-artifact detection loss,and edge loss is effective in improving the perceived quality of the restored image.