Real-Time Static Hand Gesture Recognition Using EfficientNetV2-S

This paper presents a Convolutional Neural Network (CNN) for static, frame-based hand gesture recognition under real-time constraints, using the EfficientNetV2-S architecture. EfficientNetV2-S is selected through a two-stage process. First, published EfficientNetV2 benchmark results are reviewed to assess the trade-off between accuracy, inference speed, and training cost. Second, EfficientNetV2-S is evaluated through direct experimental comparison on the HAGRIDv2 dataset with several other architectures, under identical training conditions. Hyperparameters are optimised using Bayesian search with early stopping. The final model is trained using a progressive training strategy with increasing input resolution and regularisation strength. Experimental results on selected subset of gesture classes in the HAGRIDv2 dataset show that the proposed approach achieves 99.39% classification accuracy, while maintaining low inference latency. Additional qualitative evaluation using live camera input demonstrates stable performance in practical usage conditions.