Research On Image Super-Resolu-Tion Reconstruction Technology Based On Deep Learning

In the actual application process,due to the limitation of hardware devices and interference in the transmission of image information,it is difficult for people to obtain more image details.The software algorithm can effectively reconstruct low-resolution image into high-resolution image,improve the accuracy of information,and is conducive to the development and progress of many fields.Aiming at the in-depth exploration and research of image super-resolution reconstruction technology based on deep learning,this thesis researches the contents as follows:Most of the in-depth learning reconstruction models neglect making full use of the feature information of each convolution layer,which leads to the loss of key information.Therefore,a global residual block recursive network is proposed in this thesis.Firstly,high-density residual learning and local feature fusion are introduced to build high-performance global residual blocks.Then,through feature fusion of global residual blocks and global residual learning,information interaction between residual blocks and residual blocks is established,which is conducive to obtaining more complete context feature information of images.However,this kind of network layer is too deep,which easily leads to the difficulty of network training,thus reducing the reconstruction performance of network.With the purpose to solve this problem,Shrinking is used to remove the redundant parameters of the network model,so as to improve the training efficiency.Finally,the experimental results show that the reconstruction performance of global residual recursive network is better than that of many advanced single-scale image reconstruction algorithms.The feature information among different scales is neglected in the network model for extracting image features at a single scale.It is easy to lose a lot of effective information in the mapping stage.With the purpose to solve this problem,this thesis proposes a multi-scale residual fusion depth network for image super-resolution tasks.By establishing multi-scale feature extraction,multi-scale channel fusion and high-performance recursive mapping layer,the network can be adapted to deep reconstruction networks of different scales.Experiments show that the two multi-scale reconstruction network models generated in this thesis can greatly improve the reconstruction performance of network.