Research On Image Super Resolution Reconstruction Based On Deep Learning

The process of image generation,transmission and storage are often accompanied with quality degradation,resulting in the loss of part of the image content and causing interference to different tasks.Image super-resolution reconstruction is a technology to restore low-resolution images to high-resolution images,which is of great significance in remote sensing detection,video surveillance and other fields.In recent years,the use of deep convolutional networks for image super-resolution reconstruction has become a research hotspot,but existing network has insufficient ability to reconstruct details and structures in images.Besides they have high requirements for computing resources,and it is difficult to meet the needs of practical applications.This thesis proposes two solutions to the above problems:(1)A cross-layer non-local attention mechanism is proposed,it extends the non-local attention mechanism to the relevant features of different levels of deep neural networks;Then this thesis analyzes and improves the widely used iterative up and down sampling method.Taking into account the lack of residual correction ability in the back projection process,the attention mechanism is introduced into the back projection process to capture the remote dependence between the residual and the original feature to enhance the feature.Besides the enhanced residual feature is used to correct the features of the original high-dimensional space to improve the ability of the network to generate image details.(2)In order to reduce the weight of super-resolution network,an improved method of recursive network is proposed,it can make full use of the memory information stored in the network and improving the fitting ability of the network by letting the memory information participate in the calculation of each stage.First,adaptive learning is carried out through the channel attention mechanism,mixing memory information and original input,then only simple residual connections are used for feature mining,then the remaining memory information and refined features are merged again through long-distance connections.Besides,this thesis proposes a recursive feature fusion method for the lack of representation ability of the reconstruction layer,this method accept all features in different layers from the recursive memory module and merge them successively in the same order as the generation from recursive memory module due to their phased nature.This thesis conducts experiments on the two proposed network models on four benchmark test sets,comparing objective evaluation indicators with classic methods,and comparing reconstructed images by different methods to prove that the proposed model can bring better reconstruction results.Besides,for the two network models,we performed ablation experiments by decomposing different parts to verify the improvement effect of the proposed method on image super-resolution reconstruction.