Face Feature Point Detection Based on a Novel Two-Branch Network Model

Face recognition technology is gradually maturing, and many face recognition techniques are based on accurate feature point detection, such as facial expression recognition. Previous research has focused on images with relatively small facial changes, where the detection results are relatively good. However, the robustness of feature point detection decreases for images with large facial changes, such as exaggerated expressions and occlusions. Also, data imbalance constrains the accuracy of detection. Due to many factors such as the complexity of the model and the amount of computation, it is difficult to meet the practical requirements. Therefore, accurate, efficient and compact face feature point detection becomes a challenge. Considering these three objectives, a two-branch face feature point detection model is proposed. The model has good detection accuracy in unconstrained environment and outperforms other models. Specifically, in terms of processing speed and model compactness, a lightweight network MobileNet-v2 is utilized to construct a backbone network to predict the locations of feature points. In the training stage, an auxiliary network is constructed to locate geometrically regularized feature points by calculating three Euler angles to estimate the rotation information of each sample. In this paper, a new loss function is designed to reduce the imbalance of data by adjusting the loss weights of each sample in the training set. In this study, the model is proved to be highly advantageous in all aspects through a large number of experiments, and comparative experiments are conducted on the challenging datasets of 300W, AFLW and WFLW. The experimental results are analyzed to achieve accurate, efficient and compact face feature point detection.