Abstract: In static gesture recognition, the traditional methods based on the manual feature extraction are time consuming and labor intensive, while the recognition rates are low. The existing convolution neural networks with single convolution kernel are not sufficient to extract features. To solve the problem, this paper proposes a double channel convolutional neural network, provided by two channels with diverse convolution kernel sizes. The model is able to extract image features from multiple scales and combine them in the fully connected layer. The experiments on Thomas Moeslund and Jochen Triesch gesture databases show that the model improves the accuracy of static gestures and enhances the generalization ability of convolution neural networks.

Key words: static gesture recognition, convolution neural network, double channel, convolution kernel

摘要： 针对静态手势识别任务中，传统基于人工提取特征方法耗时耗力，识别率较低，现有卷积神经网络依赖单一卷积核提取特征不够充分的问题，提出双通道卷积神经网络模型。输入手势图片通过两个相互独立的通道进行特征提取，双通道具有尺度不同的卷积核，能够提取输入图像中不同尺度的特征，然后在全连接层进行特征融合，最后经过softmax分类器进行分类。在Thomas Moeslund和Jochen Triesch手势数据库上进行实验验证，结果表明该模型提高了静态手势识别的准确率，增强了卷积神经网络的泛化能力。

关键词: 静态手势识别, 卷积神经网络, 双通道, 卷积核