Efficient Face Alignment Network For Masked Face
Yuyang Sha, Jie Zhang, Xiao Liu, Zhongqin Wu, Shiguang Shan · 2021
The sudden outbreak and global spread of the COVID-19 have severely disrupted normal life of everyone. Masks have become an effective way to suppress the transmission of the COVID-19. The widespread use of masks has brought sever challenges to many face-based algorithms, which would severely damage the accuracy of model. As the basis of various facial analysis algorithms, the performance of facial landmark detection could directly affect the precision of them. However, the existing methods cannot effectively locate facial landmarks of masked faces. In order to improve the robustness to masked face, we propose an efficient masked face alignment network named MaskFAN. Specifically, we introduce depthwise separable convolution and group operation to build a lightweight backbone. A novel loss function and data augmentation module are also used to boost the performance. Extensive experiments are conducted on masked face dataset, and the results show that our approach surpasses the state-of-the-art methods by a significant margin with small model size and limited computation cost.