基于二值化卷积神经网络的手势分类方法研究

Research on Hand Gesture Classification Based on Binary Convolutional Neural Networks

针对手势分类问题,提出了一种基于二值化卷积神经网络的手势分类方法。根据神经网络在低精度化后仍能保持较高分类准确性和鲁棒性的特点,结合传统高精度卷积网络手势分类方法与二值化方法提出一种网络结构。并通过实验研究了隐层参数对手势分类效果的影响,并与常用的方法进行了分类性能和运行效率对比。实验结果表明,所提出的方法在N=512时的表现最佳,与其他方法相比,计算效率明显提升,且错误率接近最好的结果。

A classification method based on binary convolutional neural networks has been proposed in view of some problems in current gesture classification.Based on the characteristics of neural networks,which can keep a relatively high degree of accuracy and robustness in classification even under a low precision,a proposal has been made of a new network structure with the traditional high-precision classification method of convolutional networks and the binary classification method combined together.In the process of the experiment,a research has been conducted on the effect of hidden layer parameters on the hand gesture classification,followed by a comparison between the classification performance and the operational efficiency of the conventional classification methods.The experimental results show that the proposed method has the best performance when N=512.Compared with other methods,its computational efficiency has been significantly improved,with its error rate close to the best result.