Abstract:

Aiming at the problems of traditional fire detectors with limited detection range, low accuracy of traditional fire detection algorithms and long detection time, a fire detection method YOLOV3-IMP based on improved YOLOV3 is proposed. YOLOV3 network structure is improved to enhance the ability to learn shallow features, including improvements to feature extraction networks and multi-scale detection. Initial prior frames are generated through an improved [K]-means clustering algorithm. Detection capabilities are improved for small fire areas through an improved loss function. Softer-NMS algorithm is adopted before outputting fire detection images to improve detection ability for overlapping areas. Experimental results show that the improved algorithm’s accuracy and recall rate is 91.6%, 83.2%, mAP is as high as 84.5%, and the detection speed is up to 0.28?s. It can meet the real-time and accuracy of fire detection, which can achieve effective fire detection.

Key words: fire detection, YOLOV3-IMP, multi-scale detection, Softer-NMS

摘要：

针对传统火灾探测器检测范围有限，传统火灾检测算法精度不高、检测时间长等问题，提出一种基于改进YOLOV3的火灾检测方法YOLOV3-IMP。在YOLOV3网络结构上进行改进，包含对特征提取网络改进和多尺度检测改进，提高对浅层特征的学习能力；通过改进的K-means聚类算法生成出初始先验框；通过改进的损失函数提高对小火灾区域的检测能力；在输出火灾检测图像之前采用Softer-NMS算法，提高对重叠区域的检测能力。实验结果表明，改进的算法准确率和召回率为91.6%，83.2%，[mAP]高达84.5%，检测速度可达0.28?s，可以满足火灾检测的实时性和准确性，能够实现有效的火灾检测。

关键词: 火灾检测, YOLOV3-IMP, 多尺度检测, Softer-NMS