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

Image super-resolution has always been a very popular research topic in the field of image processing.In real life,affected by the real environment and camera equipment,people often cannot obtain high-resolution images containing detailed information.Image super-resolution technology can reconstruct low-resolution images obtained due to environmental,equipment,and human factors into clear,high-resolution images that contain more detailed information.With the vigorous development of convolutional neural networks in the reconstruction of image super-resolution,the reconstruction effect of image superresolution is getting better and better.This article mainly improves the network model of the existing convolutional neural network image super-resolution algorithm.The main content of the research is mainly the following two parts.(1)Considering that most convolutional neural network models cannot make full use of the features extracted by the convolutional layer,and as the depth of the network model continues to deepen,the difficulty of training the network model increases accordingly.In response to these problems,combining the advantages of dense connections and residual networks,a multi-scale residual dense block convolutional neural network is proposed.The network model is mainly composed of feature extraction,multi-scale residual dense blocks and multi-scale sub-pixel convolution.First,the shallow information of the image is obtained through feature extraction;secondly,the extracted shallow features are extracted deeper through the multi-scale residual dense block.The residual dense block uses a combination of local dense feature fusion and local residual fusion.Ways to increase the use of the previous convolutional layer,and further perform global feature fusion of multiple residual dense blocks to increase the reusability of network feature information.At the same time,multi-scale convolution is used in the residual dense blocks to increase the network's Width,obtain the regional information of the image;then perform global residual fusion of the shallow feature information and the deep feature information to ensure the effective use of low-frequency information;finally,the image is reconstructed through multi-scale subpixel convolution,and finally high quality is obtained Reconstructed image(2)Considering that most image super-resolution reconstruction methods based on convolutional neural networks do not make full use of the inherent non-local similarity characteristics of images,and most of the current image super-resolution algorithms based on convolutional neural networks are mainly deepening Network depth and building more complex network models to learn more image features,but seldom use high-level image features.In response to these problems,combining the advantages of non-local features and second-order statistical features of images,a network structure based on the fusion of attention mechanism and multi-scale features is proposed.In the feature extraction process,the non-local features of the image and the second-order statistical features are combined to form an attention mechanism module to enhance the extraction of the structural information of the image.At the same time,the multi-scale feature module is used to extract image feature information at different scales.Get more image area information,and merge the image information extracted by the two methods,and finally reconstruct the image through sub-pixel convolution to obtain the final high-resolution image.