Improvement Of Convolutional Neural Network And Its Applied To Image Super Resolution Reconstruction

Super-resolution image reconstruction is an important technology used to improve visual effect and reconstruct high-frequency image information in the field of image processing.In recent years,due to the extensive application of in-depth learning in image reconstruction technology,which has become a research hotspot again.And Convolutional neural network is the most popular method in deep learning.According to the current research situation,there are many kinds of improved networks based on convolutional neural networks.For example: SRCNN algorithm based on extreme depth network,SRCNN algorithm based on integrating prior knowledge,SRCNN algorithm based on conditional confrontation and SRDNN algorithm,all of these methods have effectively solved various problems encountered in image reconstruction,and achieved good results in the experiment.The main content of this thesis is still based on CNN.Super-resolution image reconstruction is taken as the application background.Aiming at the problems of over-fitting of mapping function,insufficient convergence of loss function and insufficient depth of network,two improved models are proposed based on existing visual recognition algorithm and depth learning theory:(1)An improved SRCNN algorithm based on activation function and gradient descent method is proposed.In the process of image reconstruction,each pixel in the original image will have some influence on the effect of high resolution image reconstruction,and the commonly used activation functions will ignore or not effectively extract some pixels because of their own limitations.Firstly,We propose the combination of RReLUs and Softplus functions as activation functions,which improves the learning and mapping ability of activation units and effectively avoiding the over-fitting problem.Secondly,aiming at the problem that gradient descent method is easy to fall into local optimal solution and convergence speed is slow,a quadratic GDO method with additional correction coefficient is proposed,which accelerate convergence speed and avoid gradient dispersion.(2)A deep SRCNN algorithm based on feed-forward and feed-back neural networks is proposed.Firstly,we use convolution network to extract low-resolution image features,and then use de-convolution network which symmetrical with convolution network to reconstruct image features.Composite network model has the advantages of both networks.It not only has a strong learning ability for image feature information,but also can quickly reconstruct images according to the feature information learned at each level.In order to improve the training of composite model,we redefine the loss function of the network and introduce the theory of weighted Fisher criterion,so that the composite model can train the image sets with complex samples in disorder and unclassified,It improves the robustness and fault tolerance of the composite model.Finally,we validate the performance of these two algorithms through experiments,and compare them with other algorithms on Set5,Set14,image Net,Urban100 and BSD100 image sample datasets.The experimental results show that the proposed algorithm has been significantly improved in both subjective visual effects and objective evaluation indicators.