基于U-Net深度神经网络的早期火灾烟雾自动分割方法

在火灾事件监测中,为了减少数据处理量、加快探测速度,需要先分割出疑似烟雾区域。传统的烟雾分割算法大多需要设置阈值进行处理,算法的环境适应能力还需进一步提升。在研究中,使用U-Net结构的深度神经网络进行早期火灾烟雾的自动分割,通过半自动算法人工辅助分割出烟雾区域的图像样本,基于深度神经网络对分割烟雾区域进行学习,得到原始视频帧到分割结果的映射模型,并据此模型进行烟雾区域分割。在测试集上的分割实验结果表明该方法与传统方法相比,不需要设置阈值,自动化程度更高,分割速度极快,在疑似烟雾区域分割任务中性能较好。

Early smoke segmentation method based on U-Net convolutional network

In fire incident detection, in order to reduce the data processing quantity and speed up the detection, suspicious smoke region should be segmented firstly. Most traditional smoke segmentation algorithms need to set thresholds for processing, and the algorithm''s adaptability needs to be improved. In this study, deep neural network with u-net structure is used to automatically segment the early fire smoke region. Artificially semi-automatic algorithm is used to segment the smoke region. Then by learning from the sample labeled images, the mapping model from the original video frame to the segmentation result can be obtained. Smoke region segmentation is carried out according to the model. The results of the segmentation experiment on the test set show that the proposed method does not need to set any threshold, and it is fully automatic. The segmentation speed is much faster than traditional methods, and it has better performance in the suspicious smoke segmentation task.