Gesture recognition based on improved YOLOv8
he qi, honghua xu, Jiali Bao, xinli wang · 2025
Gesture recognition faces challenges such as small target sizes, blurred contours, background color similarity, and excessive parameter volume, leading to unsatisfactory accuracy and speed. To address these issues, this paper proposes a lightweight gesture recognition model called MYE (MobilenetV4-YOLOv8-ECA) based on YOLOv8. The study replaces YOLOv8's CSPDarknet with the lightweight backbone network MobileNetV4, which leverages improved neural architecture search (NAS) and distillation techniques to achieve efficient and accurate performance across various hardware platforms. To further enhance model performance, an Efficient Channel Attention (ECA) mechanism is introduced into the backbone network, constructing the MYE model. By strengthening the learning capability of useful features, this approach significantly reduces the model's parameter count and computational complexity, thereby optimizing overall gesture recognition performance. Experimental results demonstrate that compared to the original YOLOv8 model, MYE achieves a notable improvement in detection speed while reducing parameter volume and model size to 30% of the original. This refinement significantly lowers model complexity while maintaining high detection accuracy and inference efficiency.