Research On Deep Residual Super-Resolution Reconstruction Network Based On Feature Enhancement

Image Super Resolution(SR)technology is an important task in the field of computer vision,which aims to convert Low Resolution(LR)images into High Resolution(HR)images for more detailed and clear visual effects.This technology has attracted a lot of attention from the research community because it can provide powerful support for fields such as image processing and computer graphics.The improvement of image resolution is crucial for many devices such as monitors,TV sets,and cell phones,and therefore,SR technology has promising applications in many fields,such as live streaming,video surveillance,video coding and decoding,remote sensing satellite images,medical imaging,and video recovery.In recent years,with the rise of deep learning technology,the field of image SR has also achieved great success.At present,although most deep learning-based methods can improve the performance of image SR reconstruction networks by increasing the complexity of the model,there is a large amount of redundancy in the features extracted by the complex structure,and the different depth features are not effectively utilized,which results in the loss of some features during the staged feature learning process,leading to the loss of details,artifacts or texture distortion in the reconstructed images.To address these problems,the research in this thesis focuses on feature-enhanced depth-residual SR reconstruction networks based on feature enhancement,and proposes two new image SR reconstruction networks with the following main research works:(1)A adaptive feature enhancement for recurrent residual network is proposed.According to the principle of image progressive learning detail features,an adaptive feature fusion module is constructed,which can effectively integrate the shallow and deep features of images,and a cyclic mechanism is applied to this module to realize the refinement of image features.The adaptive feature fusion module mainly contains three network branches: detail attention branch,detail exploration branch,and weight assignment branch.Among them,the detail attention branch is constructed for the extraction and enhancement of shallow features,the detail exploration branch explores to learn deeper features,and the weight assignment branch learns the correlation relationship between the two branches on the channels,and adaptively assigns weights to the detail attention branch and the detail exploration branch to reduce the redundancy of features in order to better integrate the features at different depths of the image.Extensive experimental results on multiple datasets demonstrate that the results obtained by the method in this thesis are closer to the real HR images and obtain better performance metrics.(2)A two-stage feature enhancement for deep residual network is proposed.According to the two-stage learning strategy to gradually improve the feature representation and enhancement,the initial feature reconstruction block is used to reconstruct the rough HR image in the first stage,the HR image with rich detail features is reconstructed in the second stage,and finally the final reconstructed HR image is obtained by fusing the reconstructed images in the two stages.In the first stage,an attentional enhancement block is constructed to enhance the initial features and learn more detailed features by multiple convolution and deconvolution layers.In the second stage,based on the extracted features and reconstruction results in the first stage,multiple residual attention blocks consisting of spatial feature enhancement blocks and an attention interaction block are cascaded to learn correlations in feature map channels and multiple directions in space in order to further extract finer features and reconstruct fine details.Extensive experimental results demonstrate that the method in this thesis has excellent performance in both subjective and objective aspects,and outperforms some recent image SR methods in terms of parameters and metrics.