Research On Image Super-Resolution Reconstruction Algorithm And Application Based On Deep Learning

Image super-resolution technology aims to reconstruct a super-resolution image with higher resolution and carrying more plentiful and more detailed information from one or more low-resolution images.As a significant research direction in the field of image processing,super-resolution technology has important applications in the fields of satellite remote sensing,security monitoring and medical images,and has been a research hotspot in academia.In recent years,with the rapid development of deep learning in the field of computer vision,people have begun to apply deep networks to super-resolution,and utilize deep networks to learn the correlation between low-resolution images and high-resolution images in texture and geometry.And thus reconstruct a super-resolution result that is remarkably similar to the high-resolution image.However,existing deep learning-based image super-resolution reconstruction technology still has problems such as inadequate feature utilization and excessively smooth textures,and it also has certain application limitations when specific to various image types and realistic demand.In view of the above problems,the main research contents of this paper are as follows:Firstly,for single image super-resolution reconstruction,we propose a multi-scale super-resolution network based on the residual dense structure,and generates multi-scale super-resolution images in a "gradually increasing" manner.The multiplexing of features and the introduction of local and global skip connections optimize the flow of information on the network.In addition,we apply attention mechanism to the super-resolution reconstruction,so that the neural network can adaptively learn more distinguishing features by using the dependence of feature channels and spatial locations to reinforce the high-frequency details and suppress redundant information circulating in the network.Moreover,in order to generate the high-resolution detail information that has been lost in low-resolution images,we introduce the generative adversarial network,and add adversarial supervision to the existing multi-scale network to generate more convincing and more realistic textures.Secondly,in the field of remote sensing,it is difficult to obtain high-definition images due to interference from imaging conditions and equipment.Through observation,we can find the self-similarity of remote sensing images in local region,that is,we can find relatively similar blocks in the adjacent area of target block.In this paper,we propose a regional-level attention based remote sensing image super-resolution network which exploits the regional self-similar characteristics to divide features into multiple regional blocks,and learn local features on multiple regional blocks.The regional features are recombined to obtain the global features that are ultimately used for super-reconstruction.Thirdly,with the development and popularization of face recognition technology,the issue of how to protect the privacy of faces has become increasingly prominent.In this paper,we raise a novel question,namely how to simultaneously achieve privacy protection for the face area in super-resolution reconstruction?Aiming at this problem,we propose an integrated framework for image super-resolution and face privacy protection.The framework learns the face semantic information in the target image through the face attention network,and then the segmentation information will be used as update parameters of the intermediate features in super-resolution network to achieve the clarity separation of face area and non-face area,efficiently solves the potential vulnerabilities caused by image super-resolution.