Research On Improved Algorithm Of Convolutional Network In Deep Learning

Since 2006,the deep learning has attracted much attention.It is succefully used in many areas such as voice recognition,image recognition and other fields,and convolutional network based on deep learning also dominate in the field of computer vision.At present,the most effective and frequently used is the convolutional neural network(CNN),which is enhanced by the convolution structure of the unique parameters.In addition to CNN,there is a kind of simple convolution type among the convolution structures and it is the principal component analysis network(PCANet),which is characterized by only forward propagation path,being no reverse propagation of this process.Although the structure is relatively simple,it can be stacked into a deep network.The problem of deep convolution network is mainly about computationally intensive,harsh hardware conditions,large amount of parameters,slow training process,and the generalization of network is also not good enough.Convolutional network PCANet,with its simple network architecture,is well trained and performs well in recognition tasks.In this thesis,an improved method is proposed for PCANet and the application of face recognition,futtherly,a feature extraction network LPPNet or LPP_PCANet is constructed.Validity of proposed net is verified by extensive experiments conducted on multiple face datasets such as ORL,Yale and AR.Experiments show that LPPNet performs well on these datasets,and the experimental results are slightly better than PCANet,and are much better than simply using traditional face recognition algorithms Laplacianfaces.However,when there are fewer training images and more variation of facial expression,LPPNet is significantly better than PCANet.In this thesis,we study the validity of the convolution in the frequency domain by different fast algorithms in order to solve the problem of the slow training in traditional convolution process in CNN.The use of convolution theorem in the frequency domain to achieve spatial linear convolution is considered to be a very effective way.This paper proposes an unified one-dimensional FFT algorithm based on decimation-in-time splitradix-2/(2a),in which a is an arbitrary natral number.The acceleration performance of convolutional neural network is studied by using the proposed FFT algorithm on CPU environment.Experimental results on the MNIST database and Cifar-10 database show great improvement when compared to the direct linear convolution based CNN with no loss in accuracy,and the radix-2/4 FFT gets the best time savings of 38.56%and 72.01%respectively.Lastly,we propose an improved image denoising network based on low rank decomposition and Zhang's denosing network(DnCNN),which aims at solving the problem of much parameters in DnCNN.To the best of our knowledge,it is the first time to explore the network compression method for the deep learning image denoising network.In this thesis,the low rank matrix decomposition is used to compress the DnCNN successfully.It is pointed out that even if the depth of the original network(17 layers)decrease to 12,the compressed DnCNN can still achieve the promosing denosing performance.For the specific noise intensity,we can reduce the parameters of DnCNN by at least 75%,with a just slight reducetion in PSNR values while the visual effect shows no significant difference;For blind denoisng task,with no more than 0.5db reduction in PSNR,we guarantee a consistent visual representation of the standard DnCNN.