Research On Super-resolution Reconstruction Algorithm Of Noisy Image Based On Convolutional Neural Network

Due to the limitations of the hardware conditions of the image acquisition device or the random signal interference during image transmission,the acquired image may have a lower resolution or noise in the image.Therefore,the method of recovering the original image as much as possible from the low-resolution or noisy image comes into being,and the super-resolution reconstruction algorithm is the most commonly used method in this field.The existing image super-resolution reconstruction algorithm based on convolutional neural network(CNN)improves the reconstruction effect of the image to a great extent compared with the traditional method.However,as the algorithm improves,the parameter quantity and complexity of the algorithm increase,and the training period becomes longer.The reconstruction effect is still blurred,and the existing super-resolution algorithms are mostly based on lossless images,and the reconstruction effect on noise images is poor.In view of the fact that the current super-resolution reconstruction algorithm has many parameters,the training period is long,the reconstructed image is blurred,and the noise image reconstruction effect is poor.The ma:in research work of this paper is as follows:(1)For the current super-resolution reconstruction algorithm with many parameters,long training period,and reconstructed image blur.This thesis improves the image super-resolution reconstruction based on super-resolution using very deep convolutional networks(VDSR).The improved algorithm makes full use of the convolutional layer features of the neural network by increasing the number of short hopping connections between the network layers and the layers.Compared with the simple neural network algorithm that uses the cascading mode to connect the layers,it can learn more local features of images.At the same time,the number of features is extended before the nonlinear activation layer of the residual network,and the linear low-rank convolution kernel is introduced after the activation function,so that the algorithm improves the image reconstruction effect without increasing the parameters.In the algorithm,introducing weights normalization,and add a sub-pixel convolution layer at the end of the network to achieve low-resolution to high-resolution end-to-end learning,reducing the convolution operation time of each layer of the network increased by inputting high-resolution images,improving network convergence speed and reducing network training period.The experimental results show that the proposed method can speed up the convergence of the network and reduce the network training period while improving the quality of reconstructed image.(2)The effect of convolutional neural network on super-resolution reconstruction of noisy images is poor.This thesis,based on the image denoising-image and super-resolution reconstruction cascade network,due to the large difference of noise levels in different images lead to the result of the method Reconstructed image quality is poor.A multi-channel denoising network is designed to achieve image super-resolution reconstruction.The method improves the quality of the noise image super-resolution reconstructed image by reducing the noise level range of the image processed by each channel and combining the image quality evaluation algorithm designed in this thesis.The experimental results shown that the proposed method can improve the super-resolution reconstruction of images with uncertain noise levels.