Research On Image Super-Resolution Reconstruction Based On Lightweight Residual Network

With the rapid development of digital image processing technology and wireless communication technology,short video applications and streaming platforms are gradually emerging,making the increasing demand for video image quality,which can no longer be ignored in many application areas.However,due to the high cost and limitations of upgrading hardware to improve the resolution of devices,digital monitoring products often sacrifice resolution to some extent,thus ensuring that the recording device can work for a long time,stably,and with an appropriate frame rate to handle dynamic situations.In remote sensing,there is a similar situation,for example,with some compromise in spatial,spectral,and temporal resolution.In medicine,different medical image modalities can provide anatomical information about the human body structure and function.However,the limitations of resolution can have a significant impact on the physician’s diagnosis.Based on these facts,it proves that image resolution reconstruction techniques are still needed at this stage to solve the above mentioned problems.Deep learning-based image super-resolution(SR)methods mainly improve the performance of SR reconstruction by increasing the depth and complexity of the network,but the drawback of this approach is the large computational burden of the model.Although the reconstruction accuracy of existing SR methods has been greatly improved,the high complexity of the models and the large storage requirements make their practical application on mobile devices a pressing problem.To address these problems,this thesis proposes two new lightweight SR reconstruction models by focusing on the deep learning-based lightweight image SR reconstruction techniques,and the main research work is as follows:(1)An image super-resolution algorithm based on a lightweight bidirectional correction residual network is proposed.Since image degradation and reconstruction are two mutually inverse processes,a dominant correction module corresponding to image reconstruction and a return correction module corresponding to image degradation are constructed.In the dominant correction module,a feature fusion residual block is proposed for stepwise feature extraction in order to avoid information loss while reducing the computational cost of the model.In addition,a hybrid attention module based on the channel attention and spatial attention mechanisms is designed to forage for important features in different dimensions.To make the reconstruction results closer to the real situation,the dominant correction module is used to constrain the mapping space of the reconstruction function to assist the network training.This enables the model to significantly improve its performance while keeping it lightweight.Extensive experiments show that the network achieves a good balance between visual quality and computational overhead.(2)A lightweight image super-resolution algorithm based on a two-branch adaptive residual network is proposed.To effectively utilize the residual features,a two-branch adaptive residual module is designed by combining residual learning and attention mechanism,which can fully detect and utilize the information interaction of the two branches and retain the most useful feature information for SR reconstruction.The extracted features are aggregated into the feature bank in a hierarchical manner,a distillation fusion block is proposed to compress the hierarchical features to fully exploit the useful detail information,and finally the channel feature responses are recalibrated and weights are adaptively assigned to complete the feature distillation.The final experimental results on multiple datasets demonstrate that the network model proposed in this chapter requires only a relatively low computational overhead compared with other SR methods,while running efficiently and with excellent performance.