一种用于手势识别的知识共享融合方法

Knowledge Sharing Fusion Based Method for Gesture Recognition

提出一种用于实时识别的轻量模型,利用两个子网络分别提取不同类型的特征,在联合训练过程中进行信息共享,并通过知识蒸馏的方式将每个子网络的特征和预测结果融合到一个新的分类器中,用于手势识别.利用循环学习率的训练策略产生了更具泛化性的预测结果.在动态手势数据集DHG-14/28上该方法的识别准确率达到86.11%,验证了该方法的有效性.

For gesture recognition in human-computer interaction,a lightweight model for real-time recognition was proposed.The model consisted of two sub networks,each sub network was responsible for extracting different types of features,realizing a variety of information sharing in the joint training process,and fusing the features and prediction results of each sub network into a new classifier for gesture recognition by knowledge distillation.The cyclic learning rate strategy was adopted to produce more generalized prediction results.Extensive experiments were conducted on the dynamic gesture data set DHG-14/28,demonstrating that the method achieves 86.11%recognition accuracy,which verifies the effectiveness of the method.