Research On Image Super-resolution Network Based On Multi-feature Fusion

The purpose of image super-resolution is to reconstruct a corresponding highresolution image with more texture details from low-resolution images.In recent years,deep learning has been introduced into the field of image super-resolution.By continuously adjusting the structure of convolution neural network and optimizing the loss function,the quality of image restoration has been greatly improved.However,there are still many problems in the method of convolutional neural network: Single size convolution kernel can't mine multi-scale information of image.Hierarchical features can't be effectively fused.The prior information of image is not used effectively.Therefore,in this dissertation we studies the above problems and designs effective image super-resolution models on the basis of it.The main works of this dissertation are as follows:(1)Convolution neural network model usually uses single size convolution kernel to extract image features.But single size convolution kernel can not effectively be used to mine the multi-scale features of the image,which weakens the performance of image texture restoration algorithm,and at last affects the overall performance of the model.Therefore,a multi-scale feature fusion network is proposed in this dissertation.In this network,multiple sizes convolution kernels are used in parallel at each layer of the network,which can effectively expand the receptive field of the network,and improve the perception of different texture details,and also can effectively solve the problem of training and convergence difficulties caused by too deep network.The model was tested on four public test data sets.The experimental results show that the model can effectively improve the quality of image restoration.(2)At present,the research of image super-resolution model pays little attention to the effect of prior information on image restoration,and has insufficient ability to perceive different textures of images.Therefore,a multi-feature fusion network based on priori information is proposed in this dissertation.This network uses the convolution method to effectively merge the information of image segmentation in order to enhance the the ablity of texture perception.By using the attention mechanism to effectively merge the multi-sclae and multi-level information of image,the ability of feature extraction could be enhanced.Thus,the overall performance of the model is improved.The experimental results show that the network effectively improves PSNR/SSIM index and image restoration quality.