基于卷积神经网络的手势识别研究

Research on Gesture Recognition Based on Convolution Neural Network

随着人机交互技术的发展,手势动作作为一种自然、方便以及高效的交互方式受到人们的关注。因而对此从理论和程序执行的角度提出一个针对9种手势识别的卷积神经网络(Convolutional Neural Networks,CNN)模型。首先,从组成CNN的基本单元神经元开始,然后上升到神经网络,最终到反向传播算法。通过调整卷积神经网络中的参数(迭代次数、步长),观察不同参数对网络的均方误差和测试准确度的影响。实验结果表明,该模型和算法可以有效识别9种手势,识别准确率最高可达93. 33%.

With the development of human-computer interaction technology,gesture action as a natural,convenient and efficient way of interaction by people's attention,a convolution neural network model for nine kinds of gesture recognition is proposed from the point of view of theory and program execution.First,starting from the basic unit-neurons that makes up the CNN,then rising to the neural network,eventually the back propagation algorithm is obtained.By adjusting the parameters(number of iterations,steps)in the convolution neural network,the influence of different parameters on the mean square error and test accuracy of the network is observed.The experimental results show that the model and algorithm can effectively identify nine kinds of gestures,and the recognition accuracy is up to 93.33%.