DGF-M: Face Recognition Algorithm for Masked Face Scenarios

This article discusses the development of DGF-M, a face recognition AI model designed for scenarios where individuals are wearing masks. The model was developed by Didi's security product technology department to address the challenges of face recognition in masked environments, particularly in the context of the COVID-19 pandemic.

The DGF-M algorithm achieves high accuracy in both mask detection and face verification. In mask detection scenarios, it uses multi-task training and synthetic data generation to accurately identify faces with masks, achieving a detection miss rate of less than 0.1% and a mask detection accuracy of over 99.8%. For face verification with masks, the algorithm detects and locates effective facial regions, adaptively adjusts the weight of different facial areas, and extracts features from the effective regions. With a false acceptance rate of one in a thousand, the verification pass rate reaches 99.5% (compared to 73.1% before optimization), and with a false acceptance rate of one in ten thousand, the pass rate reaches 99.0%.

The article provides a detailed explanation of the face recognition process with masks, including case studies and an online workflow. The solution has been applied to offline driver mask verification, helping to identify mismatches and combat fraud during the pandemic. It is also being prepared for online driver mask recognition to enhance safety.

Didi's mask face recognition capabilities are now available on the Didi Cloud API marketplace, offering free or discounted access to the broader industry. The article also mentions Didi's achievements in face recognition, including ranking first in the Trillion Pairs face recognition evaluation competition and achieving top global rankings in the FRVT evaluation for criminal suspect recognition and cross-year criminal suspect recognition.

The article concludes by highlighting Didi's commitment to open-sourcing its technologies and contributing to the fight against the pandemic through the release of the DGF-M model.