Enhanced fire detection in new dataset: a fine-grained and multiscale-optimized architecture

Abstract

This paper presents a fine-grained and multiscale architecture for fire detection in complex scenes. The EFD-YOLOv8 framework extends YOLOv8 with a dynamic ghost convolution module for efficient feature extraction and a strip large-kernel spatial module for multiscale fire representation. A new fire dataset supports fine-grained classification across multiple fire types and scenarios, improving detection accuracy and robustness over coarse binary detection.