Multi-Scale Forest Fire Recognition Model Based on Improved YOLOv5s

The frequent occurrence of forest fires causes irreparable damage to the environment and economy. Therefore, accurate detection is particularly important. Due various shapes textures flames large variation in target scales, traditional fire methods have high false alarm rates poor adaptability, which results severe limitations. To address problem low accuracy caused by multi-scale characteristics changeable morphology fires, this paper proposes YOLOv5s-CCAB, an improved model based on YOLOv5s. Firstly, coordinate attention (CA) was added YOLOv5s order adjust network focus more features. Secondly, Contextual Transformer (CoT) introduced into backbone network, a CoT3 module built reduce number parameters while improving ability capture global dependencies images. Then, changes were made Complete-Intersection-Over-Union (CIoU) Loss function improve network’s for targets. Finally, Bi-directional Feature Pyramid Network (BiFPN) constructed at neck provide with effective fusion capability extracted experimental dataset show that YOLOv5s-CCAB increases [email protected] 6.2% 87.7%, FPS reaches 36.6. This indicates has speed. method can reference real-time, fires.