双通道卷积神经网络深度学习方法研究

发布时间：2018-07-12 21:41

本文选题：深度学习 + 卷积神经网络　；参考：《中国民航大学》2017年硕士论文

【摘要】：卷积神经网络作为深度学习的一个分支,已经在多个领域取得了巨大成功,其中神经网络的深度是其取得成功的关键。然而,越深的神经网络训练起来就越困难,因此,论文将主要研究如何设计并训练一个深度卷积神经网络模型。论文主要工作如下:首先,论文结合卷积神经网络的特点,提出了一种单通道卷积神经网络(Single-channel Convolution Neural Networks,SCNN)模型,详细介绍了该模型的实现方式和训练流程。为了降低模型过拟合的风险,将Dropout算法引入SCNN,提高SCNN的泛化能力。为了加快模型训练速度,将批归一化(Batch Normalization,BN)算法引入SCNN模型,对网络所有卷积层的激活值进行批归一化处理。然后,为解决深度卷积神经网络由于梯度消失而导致训练困难的问题,提出一种快速、高效的双通道卷积神经网络(Dual-Channel Convolution Neural Networks,DCNN)模型,该模型由两种通道构成:直通通道和卷积通道。直通通道负责保障深度网络的畅通性;卷积通道负责深度网络的学习。考虑到深层网络在训练时容易出现性能不稳定的问题,在卷积通道上引入卷积衰减因子,对其响应数据进行约束。为了保证各通道数据维度的一致性,设计了一种双池化层对同一特征图进行降采样。在CIFAR-10、CIFAR-100和MNIST 3个标准图像识别数据集上,DCNN分别取得了94.53%,73.40%和99.74%的分类准确率,无论是神经网络的可训练深度、稳定性和分类精度,DCNN都明显优于现有的大多数深度卷积神经网络模型。最后,将论文所提的DCNN模型应用于航班延误预测。在美国交通运输统计局提供的真实航班运行数据上,DCNN模型预测航班延误等级的准确率为92.08%,有关成果可以为机场和旅客提供服务和指导,具有非常重要的现实意义。
[Abstract]:As a branch of deep learning, convolutional neural networks have achieved great success in many fields, in which the depth of neural networks is the key to its success. However, the deeper the neural network is, the more difficult it is to train it. Therefore, this paper will focus on how to design and train a deep convolution neural network model. The main work of this paper is as follows: firstly, combining the characteristics of convolution neural network, a single-channel convolution neural network (SCNN) model is proposed, and its implementation and training flow are introduced in detail. In order to reduce the risk of model overfitting, Dropout algorithm is introduced into SCNN to improve the generalization ability of SCNN. In order to speed up the model training, batch Normalization BN (batch Normalization BN) algorithm is introduced into the SCNN model to normalize the activation values of all convolution layers in the network. Then, a fast and efficient Dual-Channel Convolution Neural Network (Dual-Channel Convolution Neural Network) model is proposed to solve the problem that the deep convolution neural network is difficult to train due to the disappearance of the gradient. The model consists of two channels: through channel and convolution channel. The through channel is responsible for ensuring the smooth flow of the deep network, and the convolution channel is responsible for the study of the deep network. Considering that the deep network is prone to unstable performance during training, the convolution attenuation factor is introduced into the convolution channel to constrain the response data. In order to ensure the consistency of each channel data dimension, a double cell layer is designed to de-sample the same feature map. On the CIFAR-10 CIFAR-100 and MNIST standard image recognition data sets, DCNN has achieved 94.53% and 99.74% classification accuracy, respectively. Both the training depth, stability and classification accuracy of the neural network are obviously superior to most of the existing deep convolution neural network models. Finally, the DCNN model proposed in this paper is applied to flight delay prediction. Based on the actual flight operation data provided by the United States Transportation Statistics Bureau, the accuracy of DCNN model in predicting flight delay grade is 92.08. The results can provide service and guidance for airports and passengers, which is of great practical significance.
【学位授予单位】：中国民航大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13;TP18

【参考文献】