Research On Image Super-Resolution Reconstruction Based On Feature Enhancement

Image super-resolution(SR)reconstruction has received increasing attention from the research community in recent years.Image SR reconstruction aims to restore low resolution(LR)images lacking details into high resolution(HR)images with higher visual quality and fine details.This technology can reduce the cost of hardware equipment,overcome the limitations of the shooting environment,and has broad practical application prospects in the fields of military,public security,medical diagnosis and so on.Deep learning has powerful feature learning and representation capabilities,and has achieved great success in the field of computer vision.As a branch of machine learning,deep learning aims to automatically learn the mapping relationship between input and output directly from data.This thesis mainly uses deep learning technology to build an image SR reconstruction model.The main research work of this thesis is as follows:1.In order to reconstruct more texture information and improve the visual effect of the reconstructed image,this thesis embeds the idea of texture transfer into the image SR reconstruction network,and proposes a reference-based texture enhancement network for single image SR reconstruction.First,based on the dense residual structure and attention mechanism,a preliminary reconstruction module is constructed to reconstruct the initial features of HR images.Then,in order to further enhance the feature of the reconstructed image,a multi-scale texture enhancement module is designed to transfer the rich texture of the reference HR image to the initial features of the reconstructed HR image.In addition,a mixed attention module is added to the two modules to enhance the important features of the image in channel and spatial dimensions.Finally,the effectiveness of the designed multi-scale texture enhancement module and attention module is proved by ablation experiments,and the method is compared with some advanced models in recent years.The experimental results show that the reconstruction results of the proposed network achieve higher indicators compared with the same type of referencebased reconstruction models,reconstruction results with better visual effects are obtained compared to single-image reconstruction models.2.This thesis proposes a face SR reconstruction method based on feature enhancement to solve the influence of face priors on the reconstruction effect when estimated from LR images or rough SR images.The network improves the quality of preliminary SR results by designing a feature information enhancement module to enhance the feature information extracted from the input LR images.And extract face prior information from preliminary SR results through hourglass network.In order to obtain clearer face images,a fusion module is added to effectively integrate feature information and face prior information,and finally generate enhanced face SR results with high-fidelity details.Finally,ablation experiment is designed to demonstrate the effectiveness of the enhancement module,and the proposed model is compared subjectively and objectively with several advanced face SR methods.The experimental results show that the proposed algorithm can achieve better results in subjective visual effects and objective evaluation.