Abstract: To address issues of missed detections and low accuracy in existing smoke detection algorithms when dealing with variable smoke patterns in small-scale objects and complex environments, FAR-YOLO was proposed as an enhanced smoke detection model based on YOLOv8. The model adopted Fast-C2f structure to optimize and reduce the amount of parameters. Adaptive Feature Alignment Module (AFAM) was introduced to enhance semantic information retrieval for small targets by merging and aligning features across different layers during point sampling. Besides, FAR-YOLO designed an Attention- Guided Head (AG-Head) in which feature guiding branch was built to integrate critical information of both localization and classification tasks. FAR-YOLO refines key features using Dual-Feature Refinement Attention module (DFRAM) to provide complementary guidance for the both two tasks mentioned above. Experimental results demonstrate that FAR-YOLO improves detection accuracy compared to existing. There's a 3.5% Precision increase and a 4.0% AP50 increase respectively in YOLOv8. Meanwhile, the model reduces number of parameters by 0.46M, achieving an FPS of 135, making it proper for real-time smoke detection in challenging conditions and ensuring reliable performance in various scenarios.

Yuanpan Zheng, Binbin Chen, Zeyuan Huang, Yu Zhang, Chao Wang and Xuhang Liu, “Smoke Detection Model with Adaptive Feature Alignment and Two-Channel Feature Refinement” International Journal of Advanced Computer Science and Applications(IJACSA), 16(4), 2025. http://dx.doi.org/10.14569/IJACSA.2025.0160459

@article{Zheng2025,
title = {Smoke Detection Model with Adaptive Feature Alignment and Two-Channel Feature Refinement},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2025.0160459},
url = {http://dx.doi.org/10.14569/IJACSA.2025.0160459},
year = {2025},
publisher = {The Science and Information Organization},
volume = {16},
number = {4},
author = {Yuanpan Zheng and Binbin Chen and Zeyuan Huang and Yu Zhang and Chao Wang and Xuhang Liu}
}