Research on Lightweight dynamic gesture recognition model driven by Meta-learning under Small Sample conditions

This study is dedicated to addressing the challenges of model efficiency and generalization capability in dynamic gesture recognition under small-sample conditions. It proposes a novel and efficient gesture recognition framework that integrates meta-learning strategies with a lightweight network architecture. By combining the meta-learning algorithm based on optimization with lightweight technologies such as Neural Architecture Search (NAS) and Knowledge Distillation (KD), the framework realizes rapid adaptation and accurate recognition of dynamic gestures with a small number of samples. In order to evaluate the performance of the method, systematic experiments are carried out on several standard datasets such as DHG-14, SHREC2017, and FPHA, and fewshot tasks including cross-user differences, perspective changes, background interference and other challenges are constructed. The experimental results are not only compared with traditional models such as 3D-CNN and ST-GCN, but also compared with mainstream meta-learning baselines (such as MAML and ProtoNet). Simulation results show that the proposed lightweight meta-learning model significantly reduces the model complexity and computational overhead while maintaining high recognition accuracy.