为提高卷积神经网络在火焰图像识别中的正确率、响应速度和抗误报警能力，实现对极早期火焰的有效探测和保证探测器的可靠性，针对卷积神经网络火焰识别算法进行改进，提出了一种采用双尺度去噪的深度可分离卷积火焰图像识别算法。通过对移植该算法的火焰图像探测器进行响应速度和抗误报警实验和在与某型采用神经网络算法的火焰图像探测器的对比实验中分别发现：探测器对放置在距其15 m处的5 cm×5 cm的火盘响应速度为2.8 s；探测器对强光干扰源和高热干扰源有较强的抗干扰能力；采用改进算法的探测器在灵敏度、响应速度和抗干扰能力上均优于原始算法的探测器，说明采用双尺度去噪的深度可分离卷积火焰图像识别算法的火焰图像探测器对火焰具有更高的灵敏度、响应速度和较强抗干扰能力。

In order to improve the accuracy rate, response speed and antifalse alarm ability of convolutional neural network (CNN) in the flame image recognition, realize the effective detection of very early flame and ensure the reliability of the flame image detection system, a depthseparable convolutional flame image recognition algorithm using dualscale denoising was proposed to improve the CNN flame recognition algorithm. The response speed and antifalse alarm experiments of the flame image detector transplanted with this algorithm and the comparison experiment with a flame image detector using neural network algorithm showed that the response speed of the detector to a 5 cm×5 cm fire disk placed at 15 m away from it was 2.8 seconds; the detector had strong antiinterference ability to strong light source interference and high thermal source interference; the detector using the improved algorithm was superior to the original one in sensitivity, response speed and antiinterference ability. The results show that the flame image detector using the dualscale denoising depthseparable convolutional flame image recognition algorithm has higher sensitivity, response speed and stronger antiinterference ability to the flame.

TK221