Abstract:

To accurately and efficiently detect fires in complex tunnel environments, an enhanced YOLOv8s-based tunnel fire detection algorithm was proposed. Firstly, the Cross-Stage Partial Transformer Block (CSP-PTB) module was introduced to reconstruct the backbone network structure, thereby reducing computational complexity while preserving feature extraction capabilities. Secondly, CBAM was integrated to enhance the perception of the model of key areas and improve the discriminative power of feature representation. Finally, the Normalized Wasserstein Distance (NWD) loss function was employed to optimize the training process, effectively addressing the issue of insufficient detection accuracy for small targets. Experimental results demonstrate that the improved YOLOv8s model achieves a mean average precision (mAP) of 0.848, representing a 2% improvement over the original YOLOv8s model. The recall rate reachs 0.812, marking a significant increase of 9.3% compared to the original model. Additionally, the computational cost (GFLOPS) of the improved model is reduced by 6.7%, achieving dual objectives of performance enhancement and efficiency optimization. Compared with mainstream object detection models such as Faster R-CNN(Faster Region-based Convolutional Neural Network), SSD(Single Shot MultiBox Detector), and YOLOv5s, the improved model exhibits superior performance, with mAP improvements of 7.3%, 10.1%, and 4.2%, respectively, thus meeting the stringent requirements for tunnel fire detection.

Key words: YOLOv8 model, tunnel fire detection, convolutional neural network(CNN), convolutional block attention module (CBAM), loss function