Image Super-Resolution Reconstruction Algorithm Based On Convolution Neural Network

Image super-resolution has been a hot topic in the field of image processing.Convolutional neural network(CNN)is widely used to solve the problem of super-resolution because of its rich expression ability.In this paper,the existing image super-resolution method based on convolution neural network are studied and analyzed.And it is found that many algorithms in order to get better performance at the cost of computing resources or memory.In order to adapt to the popularity of mobile terminals and we media,this paper aims to build a simple and effective lightweight super-resolution reconstruction algorithm.The research is as follows:(1)In this paper,the basic model VDSR was deeply analyzed.Firstly,the framework of model was improved,and the low resolution image was directly put into the network for training.The operation of up-sampling is completed by adding the up-sampling method to the end layer of the network.This method not only saves computing resources,but also avoids the noise that may be introduced by up-sampling in advance.In addition,group convolution was introduced in the design of network structure.Some features were divided into several groups in the channel direction,and each group was stitched after convoluting.This method can significantly reduce the amount of network parameters.Finally,the global residual was used to interpolating up-sample the input image for a target size,and added it to the residual image acquired by the network pixel by pixel.The experimental results show that,the performance improvement of this method is limited,but the amount of parameters is reduced by more than one hundred thousand.(2)In this paper,a dual attention module was added to the network to adaptively recalibrate the characteristic response in the dimensions of both channel and space,the network was focused on learning somethings that are conducive to reconstructing the characteristics of high-frequency information,and solved the problem of backward transmission without distinguishing features in deep network.The expression ability of the network is improved,and then the reconstruction performance of the algorithm is enhanced.The experimental results show that the proposed method further reduces the amount of model parameters,which is only about 1/14 of the basic model.At the same time,compared with the basic model,the peak SNR of the four test dataset is increased by 0.273 db,0.222 db and 0.285 db when the up-sampling multiple is set to 2,3 and 4 respectively.Compared with other mainstream reconstruction methods on public data sets,this paper proves that the proposed method has a more suitable balance between the amount of model parameters and reconstruction effect.