[Paper Review] RetinaMask: A Face Mask detector

Mingjie Jiang, Xinqi Fan+1 more|arXiv (Cornell University)|May 8, 2020

Face recognition and analysisComputer Science35 references156 citations

TL;DR

This paper proposes RetinaFaceMask, a one-stage face mask detector using a feature pyramid network and a novel context attention module to enhance detection accuracy. It achieves state-of-the-art performance with 2.3% and 1.5% higher precision and 11.0% and 5.9% higher recall on a public dataset, while also enabling lightweight deployment via MobileNet.

ABSTRACT

Coronavirus disease 2019 has affected the world seriously. One major protection method for people is to wear masks in public areas. Furthermore, many public service providers require customers to use the service only if they wear masks correctly. However, there are only a few research studies about face mask detection based on image analysis. In this paper, we propose RetinaFaceMask, which is a high-accuracy and efficient face mask detector. The proposed RetinaFaceMask is a one-stage detector, which consists of a feature pyramid network to fuse high-level semantic information with multiple feature maps, and a novel context attention module to focus on detecting face masks. In addition, we also propose a novel cross-class object removal algorithm to reject predictions with low confidences and the high intersection of union. Experiment results show that RetinaFaceMask achieves state-of-the-art results on a public face mask dataset with $2.3\%$ and $1.5\%$ higher than the baseline result in the face and mask detection precision, respectively, and $11.0\%$ and $5.9\%$ higher than baseline for recall. Besides, we also explore the possibility of implementing RetinaFaceMask with a light-weighted neural network MobileNet for embedded or mobile devices.

Motivation & Objective

• Address the growing need for automated face mask detection in public health and safety contexts.
• Improve detection accuracy and efficiency for face mask recognition in real-world scenarios.
• Develop a model that maintains high performance while being deployable on mobile or embedded devices.
• Overcome limitations in existing face mask detection systems by introducing context-aware feature learning and confidence-based prediction filtering.

Proposed method

• Employ a one-stage object detection framework based on RetinaNet, integrating a feature pyramid network to fuse multi-scale feature maps for improved localization.
• Introduce a novel context attention module that enhances feature representation by focusing on regions relevant to mask detection.
• Apply a cross-class object removal algorithm to filter out low-confidence predictions and reduce false positives via IoU thresholding.
• Design a lightweight variant using MobileNet as the backbone network to enable efficient inference on mobile and embedded platforms.
• Train the model end-to-end on a public face mask detection dataset with data augmentation for robustness.
• Optimize the loss function using focal loss to address class imbalance between masked and unmasked faces.

Experimental results

Research questions

• RQ1 Can a one-stage detector with feature pyramid fusion and attention mechanisms achieve superior face and mask detection performance compared to existing methods?
• RQ2 How effective is the proposed context attention module in improving detection accuracy for face mask regions?
• RQ3 To what extent does the cross-class object removal algorithm reduce false positives and improve detection reliability?
• RQ4 Can the model maintain high accuracy while being efficiently deployed on lightweight mobile or embedded hardware?

Key findings

• RetinaFaceMask achieves 2.3% higher precision in face detection and 1.5% higher in mask detection compared to the baseline model.
• The model records 11.0% higher recall for face detection and 5.9% higher for mask detection, indicating improved detection of masked faces.
• The proposed context attention module significantly enhances feature representation, particularly in complex or occluded scenarios.
• The cross-class object removal algorithm effectively reduces low-confidence predictions, improving overall detection reliability.
• The lightweight MobileNet-based variant enables efficient deployment on mobile and embedded devices without substantial accuracy loss.
• Overall, RetinaFaceMask sets a new state-of-the-art performance on the public face mask detection benchmark.