Abstract: A smoke and fire detection algorithm based on improved YOLOv5s is proposed to address the issues of low detection accuracy and long time consumption in urban building environments.Firstly,the prior boxes for the fireworks dataset are reclustered using K-means;embed CA(Coordinate Attention) mechanism in the backbone feature extraction network of YOLOv5s to suppress noise interference;drawing inspiration from the Efficient RepGFPN and BiFPN(bidirectional feature pyramid network) ideas in DAMO-YOLO,a novel neck BiGFPN is designed to reconstruct the neck of YOLOv5s and promote multi-scale fusion.In order to effectively utilize the semantic information of feature maps,a lightweight universal upsampling operator CARAFE(content aware reassembly of features) is introduced.In order to reduce the number of parameters and computation caused by model improvement,GhostNet is used to reconstruct the neck of YOLOv5s.Replacing the bounding box regression loss function CIoU with SIoU accelerates the convergence of the model and improves accuracy.Experimental results show that the improved YOLOv5s has fewer parameters and computational complexity,and the mAP50 has increased by 4.6%,which can basically meet the requirements of fireworks detection.

Key words: Pyrotechnic detection, YOLOv5s, CA, BiGFPN, CARAFE, GhostNet, SIoU