Research on Lightweight Fire Detection Algorithm Based on Improved YOLOv8
In order to improve the accuracy of fire detection in the complex circumstances,reduce the rate of false detection and missed detection,and enhance the speed of fire detection,an improved strategy based on YOLOv8 algorithm is proposed.Focused at the problem that YOLOv8 network has a deep structure and high computational complexity when dealing with complex fire images or multi-target de-tection,the backbone network of the initial model is replaced by MobileViT network,so as to build a lightweight fire detection model and improve the speed of fire detection while ensuring the detection accuracy.EMA(Efficient Multi-Scale Attention Module with Cross-Spatial Learning)is embedded in the neck structure to strengthen the semantic and spatial information of the fire feature map,ensure the integrity of channel dimensional information,and improve the accuracy of fire detection.MPDIoU loss function is employed to solve the failure problem of YOLOv8 algorithm when detecting flame targets with small size and long boundary frames.Experiments were conducted on the self-built dataset and the public dataset,and the experimental results showed that the improved algorithm has the best performance on the two data sets,and the accuracy of fire detection reaches 92.3%and 95.9%respectively.Compared with the original algorithm,the accuracy of fire detection is up to 6.3 percentage points,the average accuracy is up to 9.2 percentage points,and the mAP@0.5 is up to 8.4 percentage points.The FPS reaches more than 120 frames,and the number of parameters is only 2.0 M.As a result,the improved fire detection model can well meet the requirements of real-time detection,and different data sets are endowed with fine generalization ability and robustness.
fire detectionYOLOv8MobileViTmulti-scale attention mechanismMPDIoU
国家科技期刊平台
NSTL
万方数据
