基于改进YOLOv8的火灾早期烟雾检测算法OA

Affected by factors such as low visibility and complex spatial layouts,traditional fire alarm detection systems face significant challenges in identifying smoke during the early stages of a fire.To address this issue,this paper proposes an improved image recognition model based on the YOLOv8 algorithm,named MBS-YOLO,which is used for rapid detection of early-stage fire smoke.The dataset of real-world images is optimized through data augmentation.The SimAM attention mechanism is introduced to adjust the weights of feature maps at different scales and suppress the weights of background interference.Additionally,by integrating the weighted feature fusion mechanism and bidirectional cross-scale connections of the bidirectional feature pyramid network(BiFPN),an enhanced MA-BiFPN architecture is constructed to improve feature fusion efficiency.Experimental results demonstrate that MBS-YOLO exhibits stronger environmental adaptability on the enhanced dataset.Compared with the original YOLOv8 model,the mAP@0.5 increased by 2.6 percentage points.When compared with YOLOv5n and YOLOv10n,the mAP@0.5 improved by 3.5 percentage points and 1.1 percentage points,respectively.The final improved model has a model weight file size of only 4.4 MB and achieves a detection speed of 74.0 frames/s,satisfying the requirements of lightweight design while maintaining high detection accuracy.This significantly enhances the performance of early smoke detection and provides strong technical support for rapid fire response in complex scenarios.