Industrial and Mining Fire Detection Algorithm Based on Improved YOLO

Abstract

Fire is one of the major disasters that threaten the safety of industrial and mining enterprises. In response to the limitations of existing flame and smoke detection algorithms, which fail to meet the practical application requirements of high detection rates, low false alarm rates, and strong real-time performance, this paper proposes an industrial and mining fire detection algorithm based on the improved YOLO. First, the CFM_N module is built to more effectively capture both local and global data in the feature map. Then, the improved spatial pyramid pooling module SPPFCSPC is proposed to better extract and fuse multi-scale target features. Finally, the improved downsampling module is put forward to optimize the multi-scale fusion module and to reduce the computational complexity. Comparison experiments on self-made datasets show that the proposed algorithm obtains 91.7% mAP and 87.7% F1, which are superior to the results of YOLOv5-YOLOv8 algorithms. And this algorithm achieves accurate detection of small target flames and smoke, as well as medium and large flame and smoke targets in close and medium distances. So it can meet the real-time detection task of fire in large-scale complex industrial and mining scenes.