The Application Of A Multi-layers Pre-training Convolutional Neural Network In Image Recognition

Machine learning is the inevitable product of the development of artificial intelligence to a certain stage,and machine learning methods have emerged three major branches,including connectionism,symbolism and statistical learning.Connectionism is a machine learning method that simulates human brain learning,appearing in the mid and late 1950 s.Back Propagation algorithm was proposed to make the connectionist method prosperous,but it declined in the mid-to-late 1990 s.The reason is that as the neural network becomes complicated and the number of network layers increases,the gradient disappears and it is difficult to converge.The obtained results are largely related to the initialization parameters.Deep learning method is proposed for this problem,and access to extensive research and application,so the study of neural network initialization has important theoretical and practical significance.Among many deep learning methods,Convolutional Neural Networks have achieved great success.Convolutional neural networks are proposed specifically for image processing and later used in other areas such as text processing.Convolution,pooling and weight sharing are the core of convolutional neural networks.AutoEncoder is also a kind of depth learning method.It is a three-layer neural network,which is divided into input layer,coding layer and decoding layer,and usually uses pre-training or more complicated deep networks.Although the convolution neural network has achieved great success,but when its network structure gradually becomes complicated,there is also the problem of gradient elimination,poor convergence and long training time.In view of the above problems,this paper combines the AutoEncoder with the convolutional neural network to solve the problems such as the disappearance of the gradient and the long training time,so as to obtain better results in less time when using the convolution neural network for image recognition.In this paper,a method of initializing convolutional and full-connectivity layer parameters of convolutional neural networks by using an AutoEncoder is proposed.On the basis of this,a fast sparse control method is added to the automatic encoder in order to obtain better results.The experimental results on Minist handwritten numeral database,MIT face database and Oxford 17 dataset show that :(1)the initialized scheme in this paper can achieve better results than randomized initialized methods,and can bring better timeliness to the convolution neural network,to a certain extent,solve the problems of disappearance of gradient,long training time and difficult convergence.(2)On the Oxford 17 data,the experimental results have been significantly improved,indicating that the initialization scheme in this paper has better performance on convolutional neural networks in more complex data.(3)Further experiments found that adding sparse control to the AutoEncoder can reduce the error rate of image recognition,which verifies the effectiveness of the sparse control method proposed in this paper.(4)For the initialization scheme of this paper,the performance of complex convolution neural network is better than that of simple network,this shows that this paper aimed at convolution neural network design of the initial program,there is better timeliness in complex networks to enhance.