Research on lightweight tunnel cable fire recognition algorithm based on multi-scale features.

Currently, tunnel fire detection faces challenges such as slow response times, high false alarm rates, and poor timeliness. With the rapid development of computer vision, tunnel intelligent fire detection has received extensive attention from academia and industry. In this study, a lightweight YOLO-v5 tunnel cable fire recognition algorithm with multiscale features is proposed. By replacing the YOLO-v5 backbone network, Darknet53, with Mobilenetv3-small and integrating the SimAM attention mechanism, the lightweight and detection speed of the network was improved. Second, under the premise of retaining the feature fusion method of YOLOv5 splicing feature maps by channel, a Bi-directional Feature Pyramid Network (BiFPN) was constructed and the GIou_Loss function was introduced to enhance the network's target recognition precision. By designing tunnel cable fire experiments under different wind conditions to establish a image standard database, the accuracies and feasibilities of the model are verified. The training results show that the network's mean Average Precision (mAP 99%) improved by 0.4% and the FPS (179) improved by 46.7% compared to YOLOv5. The approach can meet the needs of tunnel cable fire detection on the precision and the speed. This method provides strong scientific and technological support for the improvement of tunnel Linear fire detection, facilitate the emergency management and therefore significantly contribute to loss prevention.