针对当前人脸关键点检测算法网络模型复杂度高、在计算资源受限时不利于部署的问题,基于知识蒸馏思想,提出了一种高精度、轻量级的人脸关键点检测算法。通过改进残差网络(ResNet50)中的Bottleneck模块并引入分组反卷积,得到轻量级的学生网络。同时提出逐像素损失函数和逐像素对损失函数,通过对齐教师网络与学生网络的输出特征图与中间特征图,将教师网络的先验知识迁移至学生网络,从而提高学生网络的检测精度。实验结果表明,本算法得到的学生网络参数量为2.81M,模型大小为10.20MB,在GTX1080显卡上的每秒传输帧数为162frame,在300W和WFLW数据集上的平均误差分别为3.60%和5.50%。

In view of the high complexity of the current facial landmark detection algorithm network model, which is not conducive to deployment on devices with limited computing resources, this paper proposes a high-precision and lightweight facial landmark detection algorithm based on the idea of knowledge distillation. This algorithm improves the Bottleneck module of residual network(ResNet50) and introduces packet deconvolution to obtain a lightweight student network. At the same time, a pixel-wise loss function and a pair-wise loss function are proposed. By aligning the output feature maps and intermediate feature maps of the teacher network and the student network, the prior knowledge of the teacher network is transferred to the student network, thereby improving the detection accuracy of the student network. Experiments show that the student network obtained by this algorithm has only 2.81M parameter amount and 10.20MB model size, the frames per second on the GTX1080 graphics card is 162frames and the normalized mean error on 300W and WFLW datasets are 3.60% and 5.50%, respectively.