基于Faster R-CNN的火灾检测方法研究

Research on Fire Detection Method Based on Faster R-CNN

摘要: 火灾早期精准检测是降低生命财产损失、实现防灾减灾目标的核心技术，而烟雾作为火灾初期最显著的视觉表征，其浓度扩散、形态演变等动态特征可有效反映火势蔓延趋势。本研究以Faster R-CNN目标检测框架为基础开展系统性优化。首先，将原框架中的ResNet-50主干网络替换为ResNeXt-50网络，提升模型的特征表达能力；其次，将原框架中的FPN替换成PAFPN，显著提升了模型对小目标烟雾的检测能力，同时保持高效推理速度。实验结果表明，改进后模型的整体检测准确率较原模型提高2.6个百分点，同时处理时间维持在0.22 ms/帧。特别值得注意的是，改进模型对不同尺度烟雾目标的检测性能有所提升，其中小目标检测准确率提升1.2个百分点，大目标提升3.8个百分点。本研究为火灾烟雾的实时监测提供了高效可靠的解决方案。

Abstract: Early and precise detection of fire is a core technology for reducing life and property losses and achieving the goal of disaster prevention and mitigation. As the most prominent visual manifestation of fire in its early stage, smoke, with its dynamic characteristics such as concentration diffusion and shape evolution, can effectively reflect the trend of fire spread. This study conducts systematic optimization based on the Faster R-CNN object detection framework. Firstly, the ResNet-50 backbone network in the original framework is replaced with the ResNeXt-50 network to enhance the model’s feature expression ability. Secondly, the FPN in the original framework is replaced with PAFPN, significantly improving the model’s detection ability for small smoke targets while maintaining efficient inference speed. Experimental results show that the overall detection accuracy of the improved model is 2.6 percentage points higher than that of the original model, and the processing time remains at 0.22 ms per frame. Notably, the detection performance of the improved model for smoke targets of different scales has been enhanced, with the detection accuracy for small targets increasing by 1.2 percentage points and that for large targets by 3.8 percentage points. This study provides an efficient and reliable solution for real-time monitoring of fire smoke.