EFSD-YOLOv5s: Effective Fire and Smoke Detection Method Based on an Improved YOLOv5s Model

Le résumé fourni par la source

Ensuring effective fire and smoke detection is critical for public safety, yet it remains challenging due to the varied appearance of flames and the semi-transparent nature of smoke in complex environments. While deep learning-based detectors offer promising performance, they often suffer from high computational costs, limiting their deployment on resource-constrained edge devices. To address these issues, this paper introduces EFSD-YOLOv5s, an optimized detection model with several key improvements on original YOLOv5s. The coordinate attention (CA) module is integrated into the neck network to enhance feature representation for irregular fire shapes and translucent smoke. In the detection head, standard convolution is replaced with spatial and channel reconstruction convolution (SCConv), which improves the capture of small and occluded objects while reducing computational redundancy. Additionally, the complete intersection over union (CIoU) loss is substituted with wise-IoU (WIoU), leading to more stable training and better localization accuracy. To enable efficient edge deployment, the pre-trained model is optimized using TensorRT with post-training quantization, significantly reducing model size and accelerating inference while maintaining accuracy. Extensive experiments on benchmark datasets and real-world edge devices demonstrate that EFSD-YOLOv5s achieves superior accuracy and efficiency compared to existing YOLO models and other state-of-the-art methods, making it highly suitable for real-time fire and smoke detection in diverse scenarios.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.