Lightweight YOLOv5 Gesture Recognition Optimization Algorithm
Fanghai Li, Xitai Na, Lei Zhang · 2023
This paper proposes a lightweight LS-YOLOv5 network to address the drawbacks of existing gesture recognition algorithm models, such as complexity and large memory usage. The aim is to achieve real-time recognition on embedded devices. Firstly, a lightweight ShuffleNetV2 network and CBRM module are added to the YOLOv5 to achieve model lightweight. The ECA attention mechanism is also incorporated to boost feature representation and enhance detection precision. Finally, the SIoU loss function is used to enable more precise target localization and prediction. By conducting experiments on a gesture dataset with 10 categories, the performance and accuracy of a gesture recognition algorithm were tested. The results showed that the LS-YOLOv5 network model reduced parameter and floating-point computation by 45.7% and 48.7%, respectively, while achieving a certain degree of accuracy improvement. The model can be deployed on edge terminals.