Image Super-resolution Reconstruction Algorithm Based On Deep Convolutional Neural Network

Super-resolution Reconstruction(SR)is a software method that reconstructs high resolution(HR)images with low resolution(LR)images.It overcomes the limitations of the hardware method,can eliminate the degradation phenomenon in the reconstruction process,and has better recovery effect.It has been applied in military,urban security,medical,financial and other scenarios,and has great potential value and practical significance.This paper summarizes the development status of image SR at home and abroad,and analyzes three image SR methods based on interpolation,reconstruction and learning.Among them,the learning-based method is the most concerned by researchers,including sparse dictionary learning,neighborhood embedded learning,and deep learning.The method based on deep learning has better performance and significant reconstruction effect.Therefore,this paper deeply studies the reconstruction model of shallow network and deep network,and proposes two improved deep convolutional neural network SR algorithms.The main work is as follows:(1)A super-resolution reconstruction algorithm based on edge correction for deep convolutional neural networks is proposed.Image super-resolution reconstruction is mainly for the recovery of high-frequency information,and most of the high-frequency details exist at the edge and texture of the image,so the algorithm proposes to learn the high-resolution edge correction factor to reduce the error of the high-frequency part of the reconstructed image.The method solves the problem of blur and distortion of the details.During the training process,the advantages and disadvantages of ReLU(Rectified Linear Units),PReLU(Parametric Rectified Linear Unit)and Softplus activation functions are fully analyzed,and a new activation function is created on the basis of this,which enhances the nonlinear expression ability of the network and improves the sharpness of the image.The experimental results show that compared with many mainstream deep learning methods,the algorithm achieves better results in both visual sense and evaluation indicators.(2)A super-resolution reconstruction algorithm based on Adam optimization for deep residual network is proposed.Based on the algorithm in Chapter 3,this paper makes further research work and improves the deep convolutional neural network reconstruction algorithm.In the integral network,the combination of global residualsand local residuals is used to reduce the burden of network.In the training process,the mini batch gradient descent method is replaced by Adam optimization method to design learning rate adaptively,reduce computing memory,accelerate the convergence speed of the loss function and global optimal can be found quickly.Considering that the number of layers is too deep,the algorithm adds a separate supervisory layer in the middle of the integral network to prevent gradient explosion/dispersion problems.The experimental results show that the algorithm is more powerful and accurate in testing on various benchmark sets such as set5,set14,B100 and Urban100,and it is closer to the original image.