Journal of Safety and Environment

Traditional flame detection methods often struggle with low feature utilization efficiency, which hampers their effectiveness in meeting the demands of modern fire safety monitoring. The You Only Look Once(YOLO) framework is an efficient deep learning-based object detection architecture that enables rapid and accurate localization and identification of flame and smoke objects within images. However, the application of the state-of-the-art YOLOv13 model for flame and smoke detection remains relatively underexplored. Consequently, addressing challenges such as significant scale variations among flame and smoke targets, as well as enhancing the practical deployment potential of detection models, continues to be a key research objective. In this study, we first perform a comparative analysis between YOLOv13 and other classic models from the YOLO series to evaluate its performance advantages specifically for flame and smoke detection. To effectively manage scale variations and enhance the model's applicability, we propose a Multi-Scale Convolutional Attention(MSCA) module. This module is integrated into YOLOv13, resulting in an improved architecture termed YOLOv13s-MSCA. Experimental results demonstrate that YOLOv13 offers a more favorable balance of detection accuracy, inference speed, and practical utility compared to other versions of YOLO. Notably, the proposed YOLOv13s-MSCA model demonstrates impressive performance on the FASD dataset: flame detection accuracy under high-precision localization(mAP75) increases by 3.5 percentage points, while overall high-precision localization/detection accuracy(mAP75) improves by 2.4 percentage points. Additionally, the overall detection accuracy(mAP50-95) rises by 1.4 percentage points. On the enhanced D-Fire dataset, smoke recall improves by 1.9 percentage points. The model also showcases an improved capability for detecting small-scale flame and smoke targets with high precision. While challenges remain in accurately identifying heavily occluded objects against complex backgrounds, the proposed model exhibits strong robustness and generalization abilities, ensuring reliable detection performance across diverse and intricate scenarios. This model holds significant potential for application in real-world fire warning systems.