Research And Application Of Image Super-resolution Technology Based On Depth Learning

Image super-resolution refers to the high resolution image recovered from a low resolution image or image sequence to meet the demands of high resolution image that people often expect in many electronic image applications.High resolution means that the image has a high density of pixels,providing more detail that is indispensable in practical applications.Although image super-resolution has made a great breakthrough along with the development of deep learning,the low level of utilization of inter-layer features is common in super-resolution task,which makes the network results unable to meet the actual demand.In this paper,based on the summary of existing methods,two novel super-resolution networks are proposed,which contain feature enhancement and feature interaction schemes to promote the network to have better performance in the super-resolution task.(1)Multi-Level Feature Aggregation and Feature Self-Guided Learning Networks for Blind Super-ResolutionIn this network,a feature self-guided learning module is proposed,which allows the feature itself to be used for computational guidance,so as to improve the adaptive learning of the feature in each stage of the whole network and enhance the expression of the feature.Most of the existing super-resolution networks only use the shallow layer features in the reconstruction process,and the low utilization level of the inter-layer features makes the reconstruction effect limited.Therefore,based on blind super-resolution unsupervised degenerate representation learning,a multi-level feature aggregation and self-guided network is proposed,which realizes the hierarchical feature guidance and efficient utilization of each feature in the reconstruction process.Feature self-guided learning module and step-by-step instruction structure work together to form a complementary trend of internal self-guided learning and external guidance aggregation,so that more details can be involved in image reconstruction,and to a certain extent,improve the use level of features in the whole super-resolution process.Experiments show that the network has achieved excellent PSNR indicators and practical results.At the same time,a large number of experiments and ablation studies have demonstrated the effectiveness of the feature self-guided learning module and the feature aggregation structure.(2)Global Feature Efficient Fusion Network for Single Image Super ResolutionThis network combines the advantages of Convolution Neural Network and Transformer technology to promote the improvement of super-resolution performance.The network proposes a symmetric self-guided residual module to achieve more expressive fusion of features between different layers.At the same time,the novel residual structure contained in the module can also provide more rich feature information for the network.In order to make better use of shallow features,the network proposes a feature mutual guidance fusion module,which can achieve deep fusion of shallow features and deep features,and provide more expressive feature information for Transformer.Through experimental comparison with other networks,the network can achieve higher PSNR and clearer visualization images.Ablation studies also demonstrate the effectiveness of symmetric self-guided residual module and feature mutual guidance fusion module.Both of these networks provide novel research approaches for image super-resolution tasks from two aspects: feature enhancement and feature interaction.The specific experiments show their excellent performance,which can provide valuable references for the follow-up research in this field.