作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (4): 253-259. doi: 10.19678/j.issn.1000-3428.0056382

• 图形图像处理 • 上一篇    下一篇

基于膨胀卷积和稠密连接的烟雾识别方法

程广涛1, 巩家昌2, 赵洪伟3   

  1. 1. 国家消防工程技术研究中心 研发部, 天津 300381;
    2. 中国刑事警察学院 声像资料检验技术系, 沈阳 110854;
    3. 应急管理部天津消防研究所 研发部, 天津 300381
  • 收稿日期:2019-10-23 修回日期:2019-11-25 出版日期:2020-04-15 发布日期:2019-11-26
  • 作者简介:程广涛(1983-),男,博士,主研方向为图像处理、深度学习、视频检测;巩家昌,讲师、博士;赵洪伟,助理研究员、硕士。
  • 基金资助:
    应急管理部天津消防研究所基科费项目(2018SJ20)。

Smoke Recognition Method Based on Dilated Convolution and Dense Connection

CHENG Guangtao1, GONG Jiachang2, ZHAO Hongwei3   

  1. 1. Research and Development Department, National Center for Fire Engineering Techonology, Tianjin 300381, China;
    2. Department of Audio-Visual Information Forensic Technology, Criminal Investigation Police University of China, Shenyang 110854, China;
    3. Research and Development Department, Tianjin Fire Research Institute of MEM, Tianjin 300381, China
  • Received:2019-10-23 Revised:2019-11-25 Online:2020-04-15 Published:2019-11-26

摘要: 为更好地提取烟雾图像的全局特征,提出一种基于膨胀卷积和稠密连接的烟雾识别方法。依次堆叠膨胀率不同的膨胀卷积,扩大卷积核的感受野,使得卷积核能够感知更广泛的烟雾图像区域,在不同膨胀卷积层之间设计稠密连接机制,促进卷积层之间的信息流通,实现烟雾图像局部特征和全局特征的融合。在此基础上,构造应用于烟雾识别的深度卷积神经网络,并在训练样本和标签的凸组合上完成训练以增强模型的泛化能力。实验结果表明,与AlexNet、VGG16等方法相比,该方法具有较好的烟雾特征表达能力,能在提高烟雾识别效果的同时,减小模型尺寸效果,其实用性较好。

关键词: 烟雾识别, 卷积神经网络, 膨胀卷积, 稠密连接, 数据增强

Abstract: In order to better extract the global features of smoke images,this paper proposes a smoke recognition method based on dilated convolution and dense connection.The method stacks in order the expansion convolutions with different expansion rates to expand the receptive field of the convolution kernel,so the convolution kernel can perceive a wider area of smoke images.The dense connection mechanism is designed between different dilated convolutional layers to promote the information exchanges between layers,and realize the fusion of local and global features of smoke images.On this basis,a deep convolutional neural network is constructed for smoke recognition,and is trained on the convex combination of training samples and labels to enhance the generalization ability of the model.Experimental results show that compared with methods such as AlexNet and VGG16,this method has better smoke feature expression performance,and can achieve more reliable smoke recognition effect with a smaller model,which proves its excellent practicability.

Key words: smoke recognition, convolutional neural network, dilated convolution, dense connection, data augumentation

中图分类号: