VMF-YOLOv5: lightweight fire detection model based on improved YOLOv5

To address the challenges of low accuracy and slow speed in detecting irregular early-stage smoke and flames, this paper proposes a lightweight VMF-YOLOv5 algorithm. First, MobileNetV3 replaces the original backbone to reduce computation and accelerate inference. Second, GSConv and the VoVGS module fuse standard and depthwise separable convolutions, maintaining accuracy while reducing parameters and improving real-time performance. Finally, the SIOU loss and Soft-NMS are introduced to enhance detection of small, irregular, and overlapping flame targets. Experiments show that, compared to the original YOLOv5, VMF-YOLOv5 reduces parameters by 58.6% and model size by 7.7M, significantly improving detection speed in complex scenes and enabling better deployment on embedded devices.