Research On Image Super-resolution Reconstruction Algorithm Based On CNN

Image is the general name of all kinds of graphics and images.In today’s society,image,as an important information carrier,is the basis of human vision,the direct reflection of natural scenery and objective substances,and an important source for human beings to understand themselves and the world.The level of image resolution directly affects the amount of information that an image can contain.The higher the resolution,the greater the information it contains and the more delicate the visual experience it gives people.Therefore,people have higher and higher requirements for image resolution.In order to solve the problem of image blurring caused by the limitation of optical equipment and complex natural scenes,image super-resolution reconstruction technology come into being.Image super-resolution reconstruction technology is a key technology in the field of computer vision,which aims to restore blurred low-resolution images into high-resolution images with rich information,and is widely used in many fields such as target recognition,data transmission,medical imaging,video reconstruction,remote sensing imaging and so on.In recent years,with the continuous development of deep learning technology,convolution neural network has achieved good results in super-resolution reconstruction of low-resolution images.The strength of convolutional neural network lies in that its multi-layer network structure can learn the deep features of input data independently,and networks at different levels can learn the features at different levels.However,in practical applications,convolutional neural network will be restricted by many factors.On the one hand,the traditional network model has weak global capture ability of lowresolution image space and channel,ignores the proportion of local features in the whole mapping process,and does not make full use of hierarchical information,which leads to the deviation of texture information and edge structure of reconstructed images from real high-resolution images.On the other hand,most super-resolution reconstruction algorithms based on convolutional neural networks have some problems,such as complex network structure,large model parameters and high resource consumption,which are not conducive to the use of portable devices.Based on this,combined with deep learning technology,the researches on image super-resolution reconstruction technology are as follows:(1)In order to capture more feature information by increasing the number of filters and deepening the number of network layers,so as to achieve a better image reconstruction effect,while ignoring the influence of different channel positions and spatial positions on the reconstruction effect,resulting in insufficient multi-level feature extraction and loss of details of low-resolution images,this thesis proposes a super-resolution reconstruction algorithm based on multi-channel perceptual residual module.The algorithm uses filter branches with different receptive fields to extract features and learn the correlation between channels with different feature layers and multi-scale receptive fields.At the same time,the information in the characteristic channel domain and the spatial domain is accurately modeled,and the characteristic map containing rich information is generated,which enhances the mapping relationship of the network.Finally,the characteristic information of different branches is step polymerization,which improves the feature utilization rate.(2)In the process of image super-resolution reconstruction,limited by the parameters of each module,the problems of slow speed and harsh convergence conditions are becoming increasingly prominent during network training,and each network branch cannot exert its maximum performance.Therefore,this thesis proposes a super-resolution reconstruction algorithm based on adaptive weight adjustment.The algorithm uses multiple adaptive modules to form a nonlinear mapping unit,and uses the attention mechanism to adaptively adjust the weight ratio of different branches to learn the deeper mapping relationship between low-resolution images and highresolution images,thus solving the problem that it is difficult to learn the long-distance information of low-resolution images.Secondly,the modules are connected by mixed residuals,which effectively fuses the low-frequency information and high-frequency information in the model.(3)In recent years,the continuous development of deep learning technology has significantly improved the performance of super-resolution reconstruction of a single image.However,to achieve good image reconstruction performance is generally accompanied by complex network structure and huge calculation parameters,which is very unfriendly to convenient devices in daily life.Therefore,this thesis proposes a lightweight image super-resolution reconstruction algorithm based on spatial feature cross-fusion.By combining multiple local feature fusion modules and feature crossenhancement modules to form a nonlinear mapping unit,the high-frequency information of the image is extracted.At the same time,the residual structure is used to learn the residual information to solve the network degradation problem and extract more accurate image features.In the network structure,the symmetrical structure is used,and the refinement features of high-frequency information are extracted by multiplication of corresponding elements,which enhances the nonlinearity of the network,and heterogeneous convolution is used instead of standard convolution to effectively reduce the parameters of the network.The final experimental results show that the proposed methods have a good performance on the commonly used data sets in the field of super-resolution reconstruction,and the PSNR values and SSIM values are improved,and the overall contour and edge structure of the reconstructed image are clearer and the visual experience is better.