Research On Image Super-resolution Reconstruction Based On Depth Residual Network

As an important way for humans to obtain information,images play an irreplaceable role in daily information circulation.Image super-resolution technology is to use various digital image processing techniques to reconstruct a high-resolution image that has a mutual mapping relationship with the low-resolution image based on the low-resolution image.The traditional image super-resolution algorithm generally restores the image from the direction of the frequency domain or the spatial domain,without considering the prior information of the image itself,the restored image will have various problems,and the traditional method has strict conditions for use.requirements and therefore have certain limitations.In recent years,with the development of machine learning technology,there have been breakthroughs in super-resolution image algorithms based on deep learning methods,but there are still some shortcomings in large-scale applications in real life.This paper studies the super-resolution reconstruction algorithm of a single image based on the deep learning method.The main work is as follows:(1)Aiming at the problems of the existing super-resolution convolutional neural network with shallow layers,insufficient feature extraction,low feature utilization,and blurred reconstructed images,a cascaded and multiple residual image super-resolution reconstruction network was proposed.First,when extracting features,multiple parallel convolution operations are used to extract different levels of feature information,and the extracted feature information is fused through a cascade structure.Then,residual learning is introduced to obtain richer feature information and promote the flow of features,and at the same time alleviate the problem of gradient disappearance caused by deeper network layers.Finally,the fusion features are nonlinearly mapped to obtain high-quality reconstructed images.The experimental results show that compared with the reconstruction results of different networks,the proposed method has a great improvement in both subjective and objective indicators.(2)When using scales of different sizes to observe local areas in an image,different levels of feature information can be obtained,that is,the content of the image has diversity at different scales.In image super-resolution,the reconstruction of pixel points often has a certain relationship with the surrounding areas of different sizes.The super-resolution network using multi-scale structure often contains multiple paths,and the size of the receptive field of each path is different.The convolutional network is used to extract the features of the target through layer-by-layer abstraction,and the larger receptive field is used to extract High-level semantic features and smaller receptive fields are used to extract the detailed features of the spatial geometry of the target,and the cascade structure is used to fuse the feature information of different levels to increase the accuracy of the network reconstruction results.The attention mechanism is used to add weight information to different features to generate an attention feature map containing key high-frequency location information,which further improves the feature expression ability of the network.Aiming at the problem that the batch regularization layer included in the coordinate attention mechanism may destroy the contrast information of the original image by performing batch regularization processing on the color information,the batch regularization layer is canceled,which reduces the calculation amount of the network and optimizes the parameter space.The multi-scale feature fusion structure and attention mechanism are combined to form a residual block with skip connection.The skip connection adds the low-frequency feature information input to the residual block and the output feature information of the residual block,so that the residual block can be used in the training process.Learning high-level feature information in the middle,reducing the computational complexity of the neural network and reducing the complexity of the network.A long skip connection is used in the whole network to output the features of the original low-resolution image input to the network to the image reconstruction process of the network,which further promotes the flow of feature information in the network and enhances the quality of the reconstructed image.The paper uses two public datasets,DIV2 K and Sun-Hays80,as training sets,Set5 and Set14 as test datasets,and adds BSD100 and Urban100 as extended test sets.The reconstructed images are reconstructed on the public datasets.From the evaluation of subjective and objective aspects,it can be seen that compared with the reconstruction results of the contrast algorithm,the method proposed in this paper has achieved higher results in terms of peak signal-to-noise ratio and structural similarity,and the reconstructed images are more detailed in terms of details.Sharper,better visuals and closer to the original image.