Research On Image Super Resolution Reconstruction Algorithm Based On Convolutional Neural Network

In recent years,with the expeditious development of information technology,the traditional super-resolution reconstruction algorithm has been unable to content the existing information demands.With the breakthrough development of deep learning,convolutional neural network has become one of the main methods of image super resolution reconstruction,relying on its excellent information extraction ability.As one of the ways of perceiving the world,images often carry abundant characteristic information.However,in the actual image data capture process,the acquired image information cannot meet the actual application requirements due to various influences such as the surrounding environment,hardware equipment and transmission media.Although the problem of acquiring high-resolution images is solved by relying on super-resolution reconstruction techniques,the solution obtained by this process is not unique and is an inverse problem.That is,an input low-resolution image often corresponds to multiple output results,mainly through the continuous optimisation of the super-resolution reconstruction model to obtain a high-definition image closer to the original.Due to its own pathological nature and high application value,it has a high research space and significance in image processing.Compared with low resolution images,high resolution images have superior density values and can offer more abundant feature information.It plays a critical role in the Internet media,video surveillance,remote sensing image and other areas.However,owing to the restriction of economy and technology,it is not ideal to improve the optical component directly from the hardware level to enhance the image resolution.Therefore,people began to focus on the software level,using various algorithms to improve the resolution of the image.In this thesis,partial limitations of the current super-resolution reconstruction algorithm are studied,and two kinds of super-resolution reconstruction means based on convolutional neural network are put forward.The research work is as follows:(1)Aiming at the problem that some image super resolution reconstruction algorithms ignore multi-scale feature information and the correlation between feature channels,this thesis designs an image reconstruction model based on a multi-scale residual network.The model is composed of multi-scale module and multiple dense residual modules in series.Among them,the multi-scale module includes the multi-scale information extraction module and the multi-scale sub-pixel convolution module,which can obtain the feature information of the bottom layer and the reconstruction layer respectively.In the dense residual module,the input of each layer is connected to the output of the previous layers to assure that the feature information is fully utilised in the network transmission.At the same time,the discriminant learning of each feature channel in the network model is carried out through the attention mechanism to ameliorate the capability of the network to reconstruct the image details.The experimental datas indicate that the network achieves good reconstruction results.(2)To address the problems that most deep neural network algorithms rely on continuously increasing depth to improve the reconstruction performance,thus introducing an excessive number of parameters and neglecting the efficiency of feature information in each layer of the network,this thesis designs an image reconstruction model based on dynamic residual network.The model consists mainly of a number of dynamic residual modules,the inner part of which is composed of two branches of the modified Res2 Net residual block intersected with dynamic convolution,respectively.In particular,The modified Res2 Net residuals can elevate the representation ability of the network model for multi-scale information,while the dynamic convolution can further heighten the overall performance of the network model under inferior computational restrictions.Furthermore,the feature information extracted from each layer of the network is different.To make better use of the feature information from each layer,a combination of hierarchical feature fusion and global residual connectivity is used for information fusion in order to achieve efficient reuse of feature information.The experimental datas indicate that the network attains superior reconstruction results in both subjective visual and objective evaluation.