Masked Face Detection with Anchor-level Attention and Local Feature

As a basic task for computer vision, face detection plays an important role in the application of face recognition. However, in real-world applications, masked face detection is still a challenging problem. In this paper, we present a novel face detection framework for masked faces. Firstly, we use the anchor-level attention mechanism to reduce the impact of complex environments and occlusion on face detection. We select the ground truth with the minims attention loss to supervise the attention layer. Besides, we depart the face features and each part corresponds to the different channel of the feature vector. By the means, the occlusions on the face can be restricted in the local part of the features. The experimental results illustrate that our model improves the accuracy of the face detection task, especially in the masked face detection. Compared to SSH, the average precision of our model has an average of 2.1%, 2.1% and 5.4% improvements on WIDER FACE easy, normal and hard validation datasets, respectively, and an average of 1.6% improvement compared to FAN on MAFA dataset.