| Deep Learning is new field of machine Learning research, raised a new wave of machine Learning, attention widely in every industry. Google Brain projects, Microsoft research automatic simultaneous interpretation system, Baidu institute is deep learning technology development testimony. With the coming of the era of big data and depth model, deep learning technology has also been widely attention and development, it brings the technology will change people’s life. With the development of the field of machine learning, in recent years the convolutional neural network research also increasingly deeper. Now convolution neural network has been widely applied to various fields, and made great achievements.Convolution neural network on the basis of artificial neural network is developed a kind of efficient recognition algorithm. Typical network structure is put forward by LeCnn, calls the LeNet-5, it contains more than one phase of the convolution and sampling process, and then connect to extract the characteristics of the input to the whole layer classification results of calculation. Convolution neural network through the process of feature extraction and feature mapping, can better learning to the invariant features in the image. Now, on the basis of the researchers in the typical LeNet5, using a variety of methods to improve the structure and properties of convolution network, thus improve the universality and the recognition of images of the network effect.In this paper, according to the characteristics of the natural scene images in the further study of the convolution of the network on the basis of the theory and research results both at home and abroad, mainly to do the following:(1) The convolution of the network training algorithm was studied, based on the algorithm analysis, debug and find the optimal initialization parameter and the adaptive network structure configuration.(2) For classification results of calculation, the use of the many regional logistic regression calculation method, improve the identification accuracy.(3)For the system design a set of universal data input interface, can take to build the input image data sets to the convolutional neural network, the training and check the result of image classification. In convolution layer(4) Local deviation Shared and unshared, two methods on data set was tested and the result analysis.(5)In hidden layer use network generalization method DROPOUT, on data set was tested and the result analysis. Through the use of the above method, versatility and convolution network performance is improved. |
