基于深度残差网络的多损失头部姿态估计

Multi-Loss Head Posture Estimation Based on Deep Residual Network

为提高真实场景下头部姿态估计的准确性,提出一种采用深度残差网络的头部姿态估计方法。将深度残差网络RestNet101作为主干网络,引入优化器提高深层卷积网络训练时的梯度稳定性,使用RGB图像并采用分类器计算交叉熵损失,同时结合回归损失预测欧拉角表示头部姿态。实验结果表明,与FAN地标检测方法和无关键点细粒度方法相比,该方法在AFLW2000数据集和BIWI数据集上的平均绝对误差值更小,分别达到5.396和2.922,在300W_LP数据集上测试精度超过95%,在真实场景下具有较好的鲁棒性。

In order to improve the accuracy of headposture estimation in real scenes,this paper proposes a head posture estimation method using deep residual network.The deep residual network RestNet101 is used as the backbone network,and the optimizer is introduced to improve the gradient stability of the deep convolution network training.The cross entropy loss is calculated by using the RGB image and the classifier.At the same time,the Euler angle is predicted combined with the regression loss to represent the head posture.Experimental results show that,compared with the FAN landmark detection method and non key point fine-grained method,the proposed method has a smaller average absolute error on AFLW2000 dataset and BIWI dataset,which reaches 5.396 and 2.922,respectively.The accuracy of the method is over 95%on 300W_LP dataset,and the method has good robustness in real scenes.