摘要
为提高烟雾识别准确率,构建模块化深度卷积神经网络,进行烟雾图像特征提取和识别。模块化结构使网络架构简单而灵活,首先利用常见的深度卷积运算设计基本模块网络结构,然后仅将模块网络与全连接层依次连接,即可构建深度卷积神经网络,使对烟雾图像的表达更加具体。利用数据增强技术扩充烟雾图像训练数据,从而缓解烟雾识别中常见的过拟合问题。实验结果表明,该方法在两个测试集上分别达到了96.56%和98.82%的检测率,验证了该方法的有效性。
In order to improve the accuracy of smoke recognition,a modular deep convolutional neural network is proposed for simultaneous features extraction and recognition of the smoke image.The modular structure makes the deep network simpler and more flexible.Firstly,the basic module network structure is designed by using the common deep convolution operation,and then the deep convolutional neual network can be constructed only by connecting the modular networks with the full connection layers in turn to enhence the more abstract and concrete expression of the smoke image.The data augumentation technology is used to expand the training smoke images,so as to alleviate the common over-fitting phenomenon in smoke recognition.The experimental results show that 96.56% and 98.82% of the detection rates are respectively achieved in the two testing data sets,which proves the effectiveness of the method.
作者
程广涛
巩家昌
李建
CHENG Guang-tao;GONG Jia-chang;LI Jian(Research and Development Center,National Center for Fire Engineering Techonology,Tianjin 300381,China;Department of Audio-Visual Information Forensic Technology,Criminal Investigation Police University of China,Shenyang 110854,China)
出处
《软件导刊》
2020年第3期83-86,共4页
Software Guide
基金
应急管理部天津消防研究所基金项目(2018SJ20)。
关键词
烟雾识别
模块网络
卷积神经网络
数据增强
smoke recognition
modular network
convolutional neural network
data augumentation