Impact of Multi-size Small Anchors and Penalized IoU on Face Mask Detection

Abstract

Wearing masks as one of the most effective ways to diminish the transmission of COVID-19 increases the demand for automatic face mask detection in all countries. Face masks belong to the small objects category in images, thus introducing the challenge of training a robust face mask detector, particularly for small object detection. Feature Pyramids derived from deep convolutional neural networks are commonly used to achieve scale-invariant object detection; however, it does not reach the same level of performance in detecting face masks as in detecting larger objects. This work proposed two methods: fully utilizing the feature map extract from the neural network by adding small multiscale anchors on the last feature map, which contains the highest resolution information. The other is to replace the standard IoU calculation with a tolerant strategy for small objects. Using these two methods, we improve the accuracy of small object detection while increasing the general average precision.