摘要
面向网页交互场景下的数字手势识别存在背景复杂度、识别计算量大等问题,提出一种基于改进的支持向量机(Supportive Vector Machine,SVM)与卷积神经网络(Convolutional Neural Network,CNN)相结合的数字手势检测与识别算法。根据复杂背景下手势提取的特点,提出一种通过肤色检验对目标图像中手部图像进行提取的方法,处理得到手部轮廓作为训练数据。由于识别计算量大,识别速度成为挑战性的问题。因此,对传统卷积神经网络进行优化,采用共享权值的稀疏连接,通过稀疏滤波器进行特征提取,降低了神经网络数量级保留CNN算法在特征提取方面的优势,并且添加SVM分类器,其最终决策函数只由少数支持向量确定,在某种意义上避免了“维数灾难”,具有分类的稳定性,最终得到数字手势识别的识别率为98.87%。通过实验对比单独使用卷积神经网络或者支持向量机算法的模型,所提方法准确率提升了2%~3%。
Aiming at the problems of background complexity and recognition speed in hand gesture recognition for webpage interaction,an improved algorithm that combined Support Vector Machine(SVM)and Convolutional Neural Network(CNN)are proposed.According to the characteristics of gesture recognition in complex background,the hand image in the target image is extracted by skin color test,and the hand contour is obtained as training data.Due to the large amount of recognition calculation,the recognition speed becomes a challenging problem.Therefore,the traditional convolutional neural network is optimized.The use of sparse connections with shared weights and feature extraction through sparse filters reduce the order of magnitude of the neural network to retain the advantages of the CNN algorithm in feature extraction.At the same time,an SVM classifier is added,and the final decision function is determined by only a few support vectors.This avoids the“curse of dimensionality”in a sense and has classification stability.The final accuracy of digital gesture recognition is 98.87%.Through experimental comparison of models that use convolutional neural networks or support vector machine algorithms alone,the accuracy of the proposed method is improved by 2%to 3%.
作者
周思昀
施水才
ZHOU Siyun;SHI Shuicai(Computer School,Beijing Information and Technology University,Beijing 100192,China)
出处
《通信技术》
2021年第4期1028-1034,共7页
Communications Technology
