引入多尺度特征图融合的人脸关键点检测网络

人脸关键点的精确检测对于人脸姿态矫正、表情识别、疲劳监测等任务具有重要意义。针对当前人脸关键点检测算法对于网络的设计越来越复杂、对计算资源的需求越来越大，网络整体速度变慢，不利于移动端部署的问题，提出了一种基于多尺度关键点热度图融合的人脸关键点检测网络。通过特征图与关键点热度图的融合使网络更多地关注关键点及其周边像素的信息，同时引入了多尺度的关键点热度图融合不断加强网络对于关键点相关信息的学习能力，使用较浅的网络结构就可实现好的检测效果。实验表明该网络在WFLW数据集上取得的检测精度和速度均达到较好的效果。

Facial landmark detection network with multi-scale feature map fusion

Accurate detection of facial landmark is of great significance for tasks such as face posture correction, expression recognition, and fatigue monitoring. For the current facial landmark detection algorithm, the design of network is more and more complex, the demand for computing resources is getting larger and larger, and the overall speed of network is slow, which is not conducive to the deployment of the mobile terminal. This paper proposed a facial landmark detection network based on multi-scale key points heat-map fusion. By integrating the feature map and the key point heat map, the network paid more attention to the key points and the information of the surrounding pixels. Multi-scale key points heat-map fusion continuously strengthened the network''s ability to learn key information, and a shallow network structure could achieve good detection results. Experiments show that the detection accuracy and speed of the WFLW(wider facial landmarks in-the-wild) dataset are state-of-art.