HafaNet: An Efficient Coarse-to-Fine Facial Landmark Detection Network

Shaun Zheng,Zhan Fang,Xiuxiu Bai,Lele Ye

doi:10.1109/access.2020.3007672

Shaun Zheng, Zhan Fang + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.3007672

Copy DOI

Abstract

Facial landmark detection can be applied in various facial analysis tasks. It is a challenging problem due to the various poses and high real-time requirements. To balance accuracy and computational efficiency, we propose an efficient coarse-to-fine network by combining heatmap regression and lightweight coordinate regression to detect facial landmarks. The heatmap regression network branch employs an efficient spatial pyramid and attention mechanism to regress to a better quality heatmap. The lightweight coordinate regression network branch introduces the local region perceptron to make the small network focus on the region of interest. Our method develops the advantages of heatmap regression and coordinate regression, which can improve landmark detection. Moreover, we propose a novel two-stage FTS loss function that can further effectively solve the outlier problem in facial landmark detection. The comprehensive experiments demonstrate the effectiveness of our approach.

Highlights

Facial landmarks are applied to detect a series of points at predefined positions on face images
These predefined points contain rich geometric semantic information, and face alignment identifies the geometric structure of the human face, which can be viewed as modeling highly structured output, it becomes the basic component for face recognition [1]–[3], face editing, and facial expression recognition
Heatmap regression is easier to perform on the regression task than coordinate regression, which can greatly improve the accuracy of landmark detection

Summary

INTRODUCTION

Facial landmarks are applied to detect a series of points at predefined positions on face images. Due to the existence of occlusion, extreme posture, extreme illumination, and other problems, some areas of the face will be greatly affected At this time, once the heatmap regression network is applied to prediction, it will lead to more serious errors when we use the coordinates of these key points, which means that the key points in occlusion areas or extreme lighting areas are shifted more seriously. S. Zheng et al.: HafaNet: An Efficient Coarse-to-Fine Facial Landmark Detection Network key point coordinate is often obtained by the argmax function. The existence of outliers in the direct coordinate regression stage will make it difficult for the network to return a good overall result To solve this problem, this paper proposes an FTS loss for facial landmark detection. The FTS loss can be based on the face pose to improve the accuracy of large-pose face samples

RELATED WORK

Findings

CONCLUSION