Application And Research Of Convolution Neural Network Based On Tensor Flow

With the advent of the era of big data, to improve the performance of computer hardware, deep learning as a new machine learning method is used for analyzing and processing these data effectively. The core idea of deep learning is to take a series of nonlinear transforms, extracting from raw data from the lower to the top, and from general to specific semantics. The convolution neural network is particularly good at extracting effective features in high-dimensional complex data structures. It is this characteristic of rich expression ability that makes the convolution neural network in image recognition and classification, target detection and localization, the man-machine game, has been widely applied in areas such as unmanned. TensorFlow is the open source deep learning platform for Google, is also the most popular machine learning framework at present. This paper deals with convolution neural networks based on TensorFlow, and on the basis of this platform, the convolution neural network model is realized, and the practical problems are solved. The details are as follows:Firstly, introduces the basic methods of deep learning, research is mainly focused on the convolution convolution and pooling layer of neural network structure, and has set up a platform TensorFlow experiment, a good understanding of the working principle and TensorFlow frame structure.Secondly, specific LeNet-5 model structure is analyzed, using two convolution with a fully connected layer building a simple convolution neural network to solve the problem of handwritten digit recognition, the improved LeNet-5 model in MNIST dataset 99.3% accuracy.Finally, the cuda-convnet model described by Alex used a number of new techniques to improve,mainly for the regularization of the weights of the L2,the image was inverted random cut and other data enhancement to create more samples in each the largest pooling layer behind the use of the LRN layer to enhance the generalization of the model. The improved convolution neural network achieves an accuracy of about 88% on more complex and rich CIFAR-10 datasets.