Fire and Smoke Detection Method in Diverse Scenarios Based on YOLOv8-FD

Targeting the challenges of insufficient accuracy, missed detections, false detections, and single scene in existing fire smoke detection, an enhanced YOLOv8 fire smoke detection method is suggested in diversified scenarios. The YOLOv8-FD model was proposed. By designing EfficientViT as the backbone network to strengthen its local feature extraction capability while maintaining the detection model's power to extract global features, more comprehensive feature content of the target of interest is obtained, solving the problem of unclear flame boundary features and efficiently improving the situation of lost and fake detections, thereby enhancing detection accuracy. A C2f_EMSC group convolution structure has been designed, which can balance the increase in parameter and computational complexity attributable to deep convolution in the Neck part of the model, and ensure more comprehensive and efficient extraction of image feature information at different levels, effectively improving the problem of long real-time response time to fires. At the same time, coupling the LSKA attention mechanism with the SPPF structure to some extent optimizes the detection difficulty problem caused by small targets in the early stage of flames, reduces computational complexity, and improves detection efficiency. Finally, the detection capability of the YOLOv8-FD pattern was validated through experiments. The experimental findings indicate that the precision of this detection approach reaches 94.2%, the mAP0.5 index reaches 95%, and the FPS can be stable at 130FPS, which can ensure the real-time recognition of fire while ensuring the accuracy of identification and reducing the occurrence of false alarms and missed positives.