Implementation and Evaluation of a Flame and Smoke Detection System based on YOLO from Still Images

Abstract

In recent years, large-scale fires have frequently occurred worldwide, causing severe damage. Early detection of flames and smoke is essential for minimizing fire-related losses. This study proposes an enhanced fire and smoke detection model based on YOLOv8 to improve detection accuracy. Specifically, part of the C2f-n module in the YOLOv8 backbone is replaced with a Swin Transformer for better multi-scale feature extraction. Deformable Convolutional Networks (DCN) replace standard convolutional layers to enhance adaptability to shape and position variations. Additionally, a Convolutional Block Attention Module (CBAM) is incorporated into the detection head to improve spatial and channel-wise attention. Data augmentation techniques, such as flipping and rotation, are also applied to enhance model generalization. The evaluation used WSDY, D-fire, and a custom dataset (5,211 images). Experimental results show that the proposed model achieved an mAP50 of 0.795, outperforming YOLOv8’s 0.744, confirming improved detection accuracy. The model demonstrated stable performance in complex backgrounds, indicating its potential for real-time applications.