文档介绍:结合Inception模块的卷积神经网络图像分类方法
该论文来源于网络,本站转载的论文均是优质论文,供学习和研究使用,文中立场与本网站无关,版权和著作权归原作者所有,如有不愿意被转载的情况,请通知我们删除已转载的信息,如果需要分享,请保留本段说明。
摘 要:针对现有卷积神经网络模型参数量大、训练时间长的问题,提出了一种结合VGG模型和Inception模块特点的网络模型。该模型通过结合两种经典模型的特点,增加网络模型的宽度和深度,使用较小的卷积核和较多的非线性激活,在减少参数量的同时增加了网络特征提取能力,同时利用全局平均池化层替代全连接层,避免全连接层参数过多容易导致的过拟合问题。在MNIST和CIFAR-10数据集上的实验结果表明,%,在CIFAR-10数据集上的准确率相比传统卷积神经网络模型提高了6%左右。
关键词:卷积神经网络; Inception模块; 全局平均池化; 卷积核; 图像分类
DOI:10. 11907/rjdk. 192501
中图分类号:TP301 文献标识码:A 文章編号:1672-7800(2020)003-0079-04
Convolutional Neural Network Image Classification Method Combined
with Inception Module
QI Guang-hua, HE Ming-xiang
(College of Computer Science and Engineering, Shandong University of Science and Technology, Qingdao 266590, China)
Abstract: Aiming at the problem that the existing convolutional neural network model has large parameters and long training time, we propose a network model combining VGG model and inception module. By combining the characteristics of the two classical models, the model increases the width and depth of the network model, uses a smaller convolution kernel and more nonlinear activation, and increases the network feature extraction ability while reducing the parameter quantity. The average pooling layer replaces the fully connected layer, avoiding the over-fitting problem that is easily caused by too many parameters of the full-connected layer. Experimental results on the MNIST and CIFAR-10 datasets show that the accuracy of this method on the MNIST dataset is %. The accuracy on the CIFAR-10 dataset is about 6% higher than the traditional convolutional neural network model.
Key Words: convolutional neural network; inception module; global average pooling; convolution kernel; image classification
0 引言
智能化信息时代的到来,使得人工智能与机器学习技术得到了飞速发展。卷积神经网络[1](Convolution Neural Network,CNN)作为深度学习最常用的模型之一,由于具有良好的特征提取能力和泛化能力,在图像处理、目标跟踪与检测、自然语言处理、场