Research on Lightweight YOLOv5 Model for Edge Devices

In the workplace, smoking not only poses a threat to the environment and workers' health but also can lead to catastrophic consequences in factories with flammable materials or precision equipment. Therefore, promptly and accurately detecting and preventing smoking behavior in factories is crucial for ensuring safe production and reducing accidents. Using edge devices for detection is a natural approach. However, smoking detection models require high accuracy due to the small size of cigarette targets and the likelihood of being obscured by complex backgrounds and other objects. Traditional object detection models often focus on improving accuracy by increasing model size, such as Faster R-CNN and SSD. However, those approach typically results in models that are difficult to deploy on edge devices or low-power devices. Deploying models on embedded devices requires lightweighting to accommodate the resource constraints of the devices. This study introduces an enhanced version of the YOLOv5 model, aiming to improve model performance on edge devices while maintaining accuracy. First, to reduce redundancy in the YOLOv5 backbone network, CSPDarknet53 is replaced with MobileNetV3, leveraging its lightweight characteristics, and its feature extraction capabilities are improved by incorporating the CBAM attention mechanism. Secondly, a pruning method is proposed to further compress the model size, optimizing its performance on edge devices. The optimized model is named MobileNetV3-YOLOv5s-CBAM (MB3-YOLOv5-CBAM), and ablation experiments confirm the contributions of each optimization. Compared to the YOLOv5 model, the pruned MB3-YOLOv5-CBAM reduces the number of parameters by approximately 60.7%, memory usage by 89.5%, and achieves an inference speed of 123 FPS, an increase of about 16.2%. The model can accurately detect cigarettes, with good real-time performance a compact model size, making it more suitable for edge devices.