Research On Image Super-resolution Reconstruction Algorithm Based On Convolutional Neural Network

With the development of science and technology,low resolution image is difficult to meet the needs of many fields,and the demand for high resolution image increases day by day.In real life,digital image is often affected by complex factors during its formation,transmission and storage,and exists in the form of lower resolution.Image super resolution reconstruction technology is a kind of underlying visual technology to obtain high resolution image through algorithm design,which can not only improve the visual perception quality of the image,but also improve the accuracy of the subsequent highlevel semantic analysis.With the rapid development of deep learning,image super-resolution reconstruction algorithms based on convolutional neural networks have been widely concerned.In view of the problems existing in current image super resolution reconstruction algorithms based on convolutional neural networks,this paper proposes a corresponding image super resolution reconstruction algorithm.The main research contents are as follows:(1)Aiming at the problems that it is difficult for deep neural networks with different filters or multiple branches to reconstruct the high frequency information of images,and that multi-level feature extraction is insufficient and the range of sensitivity field is small,a multi-level residual attention network image super-resolution reconstruction algorithm based on hollow convolution is proposed.Firstly,feature extraction of the same input image is carried out by hollow convolution with different receptive fields to obtain feature maps of different scales with rich high and low frequency information.Secondly,by constructing attention-intensive residual blocks,high-frequency information features are further obtained,and local residual connections are added to integrate multi-scale feature information between multi-channels.Finally,through the comparison test on four benchmark test sets,the results show that the proposed algorithm is superior to the current mainstream superresolution reconstruction algorithms not only in terms of objective evaluation indexes but also in terms of visual effects.(2)In view of the problem that most image super resolution algorithms based on convolutional neural network adopt the same processing mode in channels and spatial domains of different importance,which leads to the failure of concentrated utilization of important features in computing resources and the large number of network model parameters,an image super resolution reconstruction algorithm based on dynamic attention network is proposed.Firstly,it changes the existing way of equalizing attention mechanisms,and assigns dynamic learning weights to different attention mechanisms by constructing dynamic attention modules,so as to obtain high-frequency information more needed by the network and reconstruct high-quality pictures.Secondly,the double butterfly structure is constructed by feature reuse to make up for the missing feature information between different attention mechanisms.Finally,a large number of experiments show that the proposed algorithm is superior to the current mainstream super-resolution reconstruction algorithms in terms of quantitative evaluation indexes and visual effects.(3)In view of the problems that most image super-resolution reconstruction algorithms based on convolutional neural network focus on learning from a large number of external training data,ignoring the internal knowledge of the image itself and paying too much attention to local features,this paper proposes an image super-resolution reconstruction algorithm based on pyramid-like residual network.Firstly,the convolutional structure of residual graph constructed by this algorithm uses a way of pre-generating graph structure to convert extracted feature graph into the vertices of the pregenerated graph structure to form the graph structure data,so as to learn the internal topological structure of the feature itself through graph convolution.Meanwhile,residual learning is used to appropriately deepen the graph convolutional network to improve the reconstruction performance.Secondly,the pyramid-like multi-void convolution structure constructed by the algorithm avoids the defect of not covering all pixels completely by making full use of the receptive field of different sizes,and better integrates the feature information of different scales.Finally,through a large number of experiments,the proposed algorithm is significantly superior to the current mainstream super-resolution reconstruction algorithms,and has better objective and subjective measurement results.