| 摘要: |
| 森林火灾初期的烟雾检测对于预防森林火灾有着重要意义。由于长距离视频监控中通常烟雾运动速度慢、特征不明显，可靠的烟雾检测仍然是一项十分具有挑战性的任务。以林火视频监控场景为研究对象，提出了一种基于循环神经网络的森林火灾识别算法。该算法由空间特征提取网络和时序神经网络组成。空间特征提取网络通过深度学习来高效提取烟和火的静态特征。循环神经网络通过中间隐藏层的记忆内容来对空间特征进一步融合，进而挖掘烟火动态特征。该方法不仅可以从图像局部区域中提
取具有区分性的空间特征，还可以通过循环神经网络来提取时序特征。实验表明，基于循环神经网络的森林火灾识别算法能够在多种多样的场景下实现96.3% 的准确率和4.5% 的误报。 |
| 关键词: 森林火灾 烟雾检测 深度学习 循环神经网络 时序特征 |
| DOI： |
| 分类号: |
| 基金项目:中央财政林业科技推广示范项目（〔2018〕GDTK-04）、国家林业科技发展项目（林业知识产权项目- 转化应用） （KJZXZC2018006） |
|
| Forest Fire Recognition Method Based on Recurrent Neural Network |
|
CAO Yichao1, WU Zepeng2, ZHOU Yufei2, WEI Shujing2, FENG Xiaoqiang1
|
|
1.Nanjing Enbo Technology Co,Ltd,Jiangsu,China;2.Provincial Key Laboratory of Silviculture,Protection and Utilization,Guangdong Academy of Forestry,Guangzhou,China
|
| Abstract: |
| Smoke in the early stage of forest is very important to prevent forest fires. Due to the slow
moving speed and nonobvious features of smoke in long-distance video surveillance, reliable smoke detection is still a very challenging task. We present a novel forest fire recognition method based on recurrent neural network (RNN) for forest fire video monitoring scenes. This architecture consists of spatial feature extraction network
and temporal neural network. Spatial feature extraction network extracts static features of fire smoke efficiently through deep learning. Through the memory of the hidden layer, RNN can further fuse the spatial features and mine the dynamic features of fire smoke . It can not only extract distinguishing spatial features from images, but also extract temporal features by recurrent neural networks. Experiments show that this method can achieve higher accuracy and less false positives in a variety of complex scenes. |
| Key words: forest fire smoke detection deep learning RNN temporal features |
