Research On Image Super-Resolution Technology Based On Attention Convolutional Neural Network

As an algorithm of computer software,image super-resolution aims to process the fuzzy part of image,so as to recover the detailed information from the low-resolution image and reconstruct the high-resolution image.At present,image super-resolution technology based on convolutional neural network has made significant breakthrough in the reconstruction of highresolution image through independently learning the features of low-resolution image.Therefore,the methods have been widely concerned by researchers at home and abroad,and has become the focus of research in the field of image super-resolution.However,image superresolution is a highly ill-posed problem.The same high-resolution image can be reconstructed by different levels of low-resolution images.The existing super-resolution methods based convolutional network still have some key unsolved defects:(1)deep network without fully exploring the available information has very large computational complexity,but which cannot always ensure high quality image.(2)In the learning of image features,low-frequency information and high-frequency information are not distinguished,and the neglect of exploring the feature information across channels,resulting as poor reconstructed effect.This paper studies the above problems and puts forward some specific solutions.The main research contents of this paper are as follows:Firstly,the attention mechanism is introduced to direct the allocation of existing resources in favor of the most informative parts.It can give weight attention to each feature map,and adaptively adjust the interdependence among feature maps by training model,so as to enhance the discriminative ability of model for image features.Secondly,a hierarchical attention dense network model is proposed for image superresolution.Due to there is a strong correlation between the low-resolution image and the reconstructed high-resolution image,it is very important to make full use of the available information in the low-resolution image.Therefore,in the overall architecture of the model,a hierarchical dense group is designed to focus on local and global feature information,and then a deep trainable network is established.Then,the attention dense module is designed to process and learn the hierarchical features,so as to enhance the ability of the network to identify the feature information.Compared with the traditional interpolation methods and the convolutional network super-resolution methods,the proposed model has better reconstructed performance on the test datasets.Finally,a multi-channel residual attention network model for image super-resolution is proposed.The structure obtains more abundant effective information by multi-source utilization of low-frequency and high-frequency feature information,and achieves the best reconstructed effect by fusion of cross-channel feature information.Two most critical components are multisource residuals group and multi-channel attention module.The former is used to capture the global low-frequency feature information of remote space and the high-frequency feature information of local network.The latter combines the features of many different channels and adaptively adjust them by using the attention mechanism,so as to the channel features have stronger discriminant representation.The validity of the proposed model is verified by experiments on datasets.Compared with previous advanced methods based on convolutional neural network,the proposed method has better effect of image super-resolution.