基于迁移学习的人脸姿态分类方法

Face Pose Classification Method Based on Transfer Learning

中文关键词:  人脸姿态分类  卷积神经网络  特征提取  迁移学习

英文关键词:face pose classification  convolutional neural network  feature extraction  transfer learning

中文摘要:

      人脸姿态分类在智能人机交互、虚拟现实、智能控制以及人脸识别等多个领域都有广泛的应用。由于人脸姿态分类过程中存在不同角度间特征重叠率高的问题，导致其分类精度过低。为提高人脸姿态分类的准确率与鲁棒性，提出了基于迁移学习的人脸姿态分类方法。该方法利用卷积神经网络的特征提取和学习能力，对特征进行识别和分类，从而得到单方向人脸姿态的训练参数。利用迁移学习，将卷积神经网络训练好的参数应用于训练两个方向的人脸姿态模型中。使用该方法在CAS-PEAL数据集上进行了实验，最终结果的准确率达到98.7%，并且与AlexNet、VGGNet和ResNet等网络模型做对比实验，得到了更好的人脸姿态分类效果。实验结果表明，所提出的方法显著提高了人脸姿态分类的准确率与鲁棒性。

英文摘要:

      Face pose classification has a wide range of applications in intelligent human-computer interaction,virtual reality,intelligent control and face recognition.Because of the high overlapping rate of features among different angles,the classification accuracy is too low.In order to improve the accuracy and robustness of face pose classification,a method based on migration learning is proposed.The feature extraction and learning ability of convolutional neural networks were used to identify and classify features,so as to obtain the training parameters of one-way face pose.Using transfer learning,the parameters trained by the convolutional neural network were applied to the training of the face pose models in two directions.To verify the effectiveness of convolutional neural networks,experiments were performed on the CAS-PEAL dataset,and the accuracy of the final result reached 98.7%.Compared with the models such as AlexNet,VGGNet and ResNet,better face pose classification results were obtained.The final results show that the proposed method significantly improves the accuracy and robustness of face pose classification.