面向森林火灾烟雾识别的深度信念卷积网络  

DBN⁃CNN for forest fire smoke recognition

作　　者：杜嘉欣 常青 刘鑫[2] DU Jiaxin;CHANG Qing;LIU Xin(College of Information and Computer,Taiyuan University of Technology,Taiyuan 030024,China;Network Management Center Wireless Room,Unit 32152 of PLA,Shijiazhuang 050000,China)

机构地区：[1]太原理工大学信息与计算机学院,山西太原030024 [2]中国人民解放军32152部队网管中心无线室,河北石家庄050000

出　　处：《现代电子技术》2020年第13期44-48,共5页Modern Electronics Technique

基　　金：海南省高新重点研发计划(ZDYF2018011)资助项目;国家自然科学基金资助项目(61828601);山西省自然科学基金资助项目(201801D121141)。

摘　　要：对于CNN的图像识别,采用随机初始化网络权值的方法很容易收敛达到局部最优值。针对林火中的烟雾图像识别,提出一种结合无监督和有监督学习的网络权值预训练算法。首先通过使用DBN预学习得到的特征初始化CNN的权值;然后通过卷积、池化等操作,提取训练样本的特征,并采用全连接网络对特征进行分类;最后计算分类损失函数并优化网络参数。实验的训练结果显示,基于DBN-CNN的森林火灾烟雾识别的准确率达到了98.5%,相比于其他算法其准确率更高。For image recognition of convolutional neural networks(CNN),the method of randomly initializing network weights can easily converge to local optimal values.In order to realize the smoke image recognition of forest fires,a network weight pre⁃training algorithm combining unsupervised and supervised learning is proposed in this paper.The weight of CNN is initialized by using the features obtained by the deep belief network(DBN)pre⁃learning.Then,the features of the training samples are extracted by means of the convolution,pooling and other operations,and the extracted features are classified by the fully connected network.Finally,the classification loss function is calculated and the network parameters are optimized.The experimental training results show that the accuracy of forest fire smoke recognition based on DBN⁃CNN reaches 98.5%,which is higher than that of other algorithms.

关 键 词：深度信念网络 森林火灾监控 烟雾识别 权值初始化 特征提取 特征分类 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP183[电子电信—信息与通信工程]