Research On Super Resolution Based On Convolutional Neural Network

High-quality,high-resolution images play an important role in modern society.The acquisition of high-resolution images is mainly considered from two approaches: hardware(improving hardware equipment directly)and software(using image super-resolution algorithm).Compared with directly upgrading the hardware to increase the image resolution and improve the image quality,the use of image super-resolution algorithm can easily and flexibly calculate the corresponding high-resolution image based on the existing low-resolution image.In addition,only improving the hardware equipment can not improve the existing lowresolution images.However,due to the limitations of the previous hardware technology level,there are a large number of low-resolution images or videos on the Internet,and it is also very meaningful to perform super-resolution restoration on these images and videos.With the development of deep learning,image super-resolution methods based on convolutional neural networks have performed well in terms of the visual quality of the reconstructed high-resolution images and their quantitative indicators,and have gradually become a hot research topic.The work in this paper focuses on image super-resolution methods based on convolutional neural networks.The main work is as follows:(1)The related concepts and algorithms are systematically described,such as lowresolution image degradation models,interpolation algorithms,learnable upsampling layers,different convolution operations,etc.And the current development status of image superresolution algorithms is elaborated.A detailed calculation formula is given for the commonly used image quality evaluation indicators.(2)An image super-resolution method based on multi-scale feature aggregation network(MFAN)is proposed.Existing super-resolution methods based on convolutional neural networks require too many parameters and very deep network structures for high-quality reconstruction,which directly lead to higher requirements for computing resources and memory storage,and are difficult to apply to resource-constrained applications.limited equipment.Aiming at this problem,this paper designs a nested multi-path structure to improve the multiscale feature extraction capability of the network.The structure consists of global dual-path and multi-scale feature extraction modules.The global dual path is used to transfer the feature information of the low-resolution space and the medium-resolution space,respectively.A multiscale feature extraction module is embedded in each global path for image multi-scale feature extraction.In addition,this paper designs an inter-scale feature projection module based on the back-projection mechanism to achieve the fusion of multi-scale feature information from different global paths.Experiments on multiple datasets such as B100 and Urban100 demonstrate the effectiveness of the proposed method.(3)An image super-resolution method(ESRNet)based on auxiliary edge optimization is proposed.The reconstruction results of image super-resolution methods based on convolutional networks are still blurred at the edge of the image,which is not ideal.This phenomenon is more pronounced in the case of high upsampling factors.Some methods explicitly reconstruct edge information,distinguish between edge semantics and non-edge semantics,and use edge information as constraints to achieve high-quality reconstruction.However,at present,such methods generally extract edge features from low-resolution images or reconstructed highresolution images in series or in parallel,and cannot make full use of low-level features in the network.In response to the above problems,this paper proposes a structure in which the edge reconstruction sub-network shares low-level features with the RGB image reconstruction subnetwork to improve the network efficiency.In order to effectively fuse edge information and reconstructed RGB images,this paper constructs a feature fusion module combined with pixel attention mechanism.In the loss function,this paper introduces edge attention loss,which suppresses the smoothing of reconstruction results.