Modifications And Applications Of Deep Convolutional Neural Networks

This dissertation is a study on image recovery and recognition arising from image processing,We construct a simple but effective convolution neural network based on the theory of deep learning and improve the relative algorithms.With the modified algorithms,we can deal with the problems of pattern recognition and image denoising.In this dissertation,we have done some research about the improvement and applications of convolution neural networks.Our research includes three parts: an image denoising algorithm based on composite convolutional neural network,a novel deep learning algorithm for incomplete face recognition: low-rank-recovery network,and matrix completion via composite convolutional neural network.The main contents are as follows:(1)Image denoising algorithm based on composite convolutional neural network.According to the theory of deep learning,the process of image denoising can be regarded as a fitting process of a neural network.In this paper,an image denoising algorithm based on composite convolutional neural network is proposed through constructing a simple and efficient composite convolutional neural network.The first stage includes two convolutional neural networks with two layers.Some initial convolutional kernels of convolutional neural network with three layers in the second stage are trained by these two networks,respectively.The training time in the second stage is decreased and the robustness of the network is enhanced.Finally,the learned convolutional neural network in the second stage is applied to denoise a new image with noises.Experimental results show that the proposed algorithm is comparable to the state of the art image denoising algorithms in peak signal to noise ratio(PNSR),structure similarity,and root mean square error(RMSE).Especially,when the noises get heavier,the proposed algorithm performs better with less training time.(2)A novel deep learning algorithm for incomplete face recognition:low-rank-recovery network.The proposed LRRNet first recovers the incomplete face images via an approach of matrix completion with the truncated nuclear norm regularization solution,and then extracts some lowrank parts of the recovered images as the filters.With these filters,some important features are obtained by means of the binalization and histogram algorithms.Finally,these features are classified with the classical support vector machines(SVMs).The proposed LRRNet performs well and efficiently for the images with heavily corrupted,especially in the case of large databases.Extensive experiments on several benchmark databases demonstrate that the proposed LRRNet possesses a rather surprise performance than some other excellent robust face recognition methods.(3)Matrix Completion via composite convolutional neural network.The main task is to recover the missing information of an incomplete matrix.For the incomplete matrix,what we do is to estimate the missing information according to the known information.So we build a map between the missing part and the known part by means of the composite convolutional neural network.With the map,we can obtain the missing information.What's more,with experiments on benchmark images,the validity and the robustness of our algorithm are provided.