Self-restrained triplet loss for accurate masked face recognition

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Face recognition is a key biometric technology, valued for its contactless and accurate nature.
• The COVID-19 pandemic necessitated widespread face mask usage, posing a significant challenge to existing face recognition systems due to occlusion.
• Developing robust face recognition solutions that perform accurately with masked faces is crucial for security and identification.

Purpose of the Study:

• To propose a novel solution to enhance the performance of face recognition systems when individuals wear face masks.
• To introduce the Embedding Unmasking Model (EUM) and a new loss function, the Self-restrained Triplet (SRT), for masked face recognition.
• To validate the effectiveness of the proposed approach across various face recognition models and datasets.

Main Methods:

• Developed the Embedding Unmasking Model (EUM) to operate on top of existing face recognition architectures.
• Introduced a novel loss function, the Self-restrained Triplet (SRT), designed to generate embeddings comparable to unmasked faces.
• Evaluated the EUM and SRT on three distinct face recognition models and four datasets (two real, two synthetic masked face datasets).

Main Results:

• The proposed Embedding Unmasking Model (EUM) with the Self-restrained Triplet (SRT) loss function significantly improved masked face recognition performance.
• The approach demonstrated effectiveness across multiple established face recognition models.
• Consistent performance enhancements were observed on both real-world and synthetically generated masked face datasets.

Conclusions:

• The EUM and SRT offer a robust solution for improving face recognition accuracy in the presence of face masks.
• This method effectively mitigates the challenges posed by mask occlusion in biometric identification.
• The proposed approach represents a significant advancement in developing reliable face recognition systems for post-pandemic scenarios.