Real-time Efficient Facial Landmark Detection Algorithms

Abstract

Lightweight models, high accuracy and real-time performance are essential for facial landmark detection algorithms. Considering these three aspects, this paper proposes a real-time and efficient face landmark algorithm. First, mobilenetV2 is used as the backbone network. Next, the traditional convolution operation is replaced with deeply separable convolution, and the shallow and deep feature maps are merged to enhance the context connection. Then multi-scale fusion output is used in the output layer to enhance the detection efficiency of small-sized faces. Finally, the Euler angle weights are introduced into the loss function, and the 14 key points in the average face model are compared with the predicted key points. During the training process, this paper proposes rotated the 300W and AFLW datasets in multi-angle to occlude the dataset and enhance the generalization ability of the model. The experimental results show that the proposed algorithm in this paper can achieve real-time and efficient facial landmark detection.