基于双分类器的自适应单双手手势识别

针对传统卷积神经网络(CNN)中仅有对单手手势语义进行识别的算法和深度学习手势识别算法中CNN的收敛性差和识别精度低的问题,提出了一种基于两个分类器的自适应单双手手势识别算法以对单手和双手进行识别。该算法的核心是联合两个分类器进行单双手手势识别。首先,采用手数分类器对手势进行分割分组预测,将手势识别转化成部分手势图像识别;其次,采用自适应增强卷积神经网络(AE-CNN)进行手势识别,利用自适应模块分析出现识别误差的原因和反馈模式;最后,在迭代次数和识别结果的基础上进行参数更新。实验结果表明,手数分类器进行手势预测分组的正确概率为98.82%,AE-CNN的收敛性优于CNN和CNN+Dropout,对单手手势的识别率高达97.87%,对基于LSP数据集自建的9类单手手势和10类双手手势的整体模型识别率为97.10%,对复杂背景和不同光照强度下手势的平均识别率为94.00%,并且具有一定的鲁棒性。

Adaptive One-Hand and Two-Hand Gesture Recognition Based on Double Classifiers

Aiming at the problem that the traditional convolutional neural network(CNN)algorithms only recognize semantics of one-hand gestures and the problems of the poor convergence and low recognition accuracy of the deep learning gesture recognition algorithm,an adaptive one-hand and two-hand gesture recognition algorithm based on double classifiers is proposed to recognize single-hand and two-hand gestures.The core of the algorithm is combining two classifiers for single-hand and two-hand gesture recognition.First,the hand number classifier is used to segment and group the gestures,and the gesture recognition is converted into partial gesture image recognition.Second,the adaptive enhanced convolutional neural network(AE-CNN)is used for gesture recognition,and the adaptive module analyzes the cause of the recognition error and feedback mode.Finally,the parameters are updated based on the number of iterations and recognition results.Experimental results show that the correct probability of the hand number classifier for gesture prediction grouping is 98.82%,the convergence of AE-CNN is better than that of CNN and CNN+Dropout,and the recognition rate of one-hand gestures is as high as 97.87%.The overall model recognition rate of 9 types of single-hand gestures and 10 types of double-hand gestures built based on LSP dataset is 97.10%,and the average recognition rate of gestures under complex backgrounds and different light intensities is 94.00%.The proposed algorithm has certain robustness.