基于TensorFlow的卷积神经网络的应用研究

发布时间：2018-09-07 10:04

【摘要】：随着大数据时代的到来,计算机硬件性能的不断提升,深度学习作为新兴的机器学习方法被用于有效地分析和处理这些数据。深度学习的核心思想是采用一系列的非线性变换,从原始数据中提取由低层到高层、由一般到特定语义的特征。而卷积神经网络尤其擅长在高维复杂数据结构中提取有效特征。正是这种丰富的特征表达能力使得卷积神经网络在图像识别与分类、目标检测与定位、人机博弈、无人驾驶等领域应用广泛。TensorFlow是谷歌公司开源的深度学习平台,也目前最受欢迎的机器学习框架。本文基于TensorFlow研究卷积神经网络,并在此平台基础之上实现卷积神经网络模型,解决实际问题。具体工作如下:首先,对深度学习的基本方法进行了介绍,重点研究了卷积神经网络结构中的卷积层和池化层,并且搭建了TensorFlow实验平台,深刻理解TensorFlow的工作原理及框架结构。其次,具体分析了 LeNet-5模型结构,使用两个卷积层加一个全连接层构建一个简单的卷积神经网络解决手写体数字识别问题,改进后的LeNet-5模型在MNIST数据集上取得99.3%的准确率。最后,对Alex描述的cuda-convnet模型使用了一些新的技巧进行改进,主要是对weights进行了 L2的正则化、对图片进行了翻转随机剪裁等数据增强以制造更多的样本、在每个卷积-最大池化层后面使用了 LRN层以增强模型的泛化能力。改进后的卷积神经网络在更复杂更丰富的CIFAR-10数据集上取得约88%的准确率。
[Abstract]:With the arrival of big data era and the continuous improvement of computer hardware performance, depth learning as a new machine learning method is used to analyze and process these data effectively. The core idea of depth learning is to use a series of nonlinear transformations to extract features from lower level to higher level and from general to specific semantics from the original data. Convolutional neural networks are especially good at extracting effective features from high-dimensional complex data structures. It is this rich feature expression ability that makes convolutional neural network widely used in image recognition and classification, target detection and location, man-machine game, driverless and other fields. Tensor flow is Google's open source in-depth learning platform. Also currently the most popular machine learning framework. In this paper, the convolution neural network is studied based on TensorFlow, and the model of convolutional neural network is implemented on this platform to solve the practical problems. The main work is as follows: firstly, the basic method of deep learning is introduced, and the convolution layer and pool layer in the network structure of convolutional neural network are studied, and the TensorFlow experimental platform is built to deeply understand the working principle and frame structure of TensorFlow. Secondly, the structure of LeNet-5 model is analyzed in detail. A simple convolution neural network is constructed by using two convolution layers and a full join layer to solve the problem of handwritten digit recognition. The improved LeNet-5 model achieves 99.3% accuracy on MNIST data set. Finally, the cuda-convnet model described by Alex is improved with some new techniques, mainly the regularization of L2 for weights and the enhancement of image data such as flipping random clipping to create more samples. The LRN layer is used after each convolution-maximum pool layer to enhance the generalization of the model. The improved convolution neural network achieves an accuracy of about 88% on the more complex and abundant CIFAR-10 datasets.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】