Research On Image Super-resolution Algorithm Based On Multi-scale Convolution Neural Network

Image super-resolution reconstruction aims to reconstruct high-resolution images from low-resolution images.In recent years,deep learning has been widely used in the field of image processing and has achieved certain results.In order to improve the performance of the super-resolution algorithm,this paper applies the multi-scale convolutional neural network to the image super-resolution problem.The main research contents are as follows:Firstly,based on the gradient information of the image,this paper proposes an image super-resolution algorithm based on gradient prior learning.This algorithm uses a multi-scale residual network to learn the difference image of high-resolution images as gradient regularization term,and then reconstructs high-resolution images by solving optimization models.Experimental results show that the proposed algorithm effectively improves the resolution of low-resolution images,and the convergence speed of the network is faster.Secondly,in order to further improve the performance of the super-resolution algorithm,this paper uses the idea of multi-scale and residual training,combined with the shrinkage-extension network structure.An image super-resolution algorithm based on multi-scale convolutional neural network and residual training is proposed.This algorithm exploits the multi-scale convolution kernels and the shrinkage-extension structure to extract image multi-scale information.Skip connection is used in the network structure to improve the quality of image reconstruction,which can transmit information effectively.Moreover,it can compensate for the loss of image details resulting from the use of down-sampling and up-sampling.Experimental result shows that this algorithm can obtain reconstructed images with significantly improved quality.Finally,in order to make full use of the prior information of the image,this paper uses wavelet transform to obtain high-low-frequency separated image information as the input of the network,and proposes a super-resolution algorithm based on undecimated wavelet transform and convolutional neural network.The algorithm uses the four sub-bands obtained from the low-resolution image one-level undecimated wavelet transform as the network input,and reconstructs high-resolution images through a deep residual network with multi-scale convolution layer.Comparison with other algorithms shows that this algorithm can achieve higher performance.