HQ2CL: A High-Quality Class Center Learning System for Deep Face Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Margin-based losses have significantly advanced face recognition by increasing inter-identity discriminability.
• Effective face recognition relies on well-defined class centers, ideally distant from each other with compact intra-class samples.
• Current methods are hindered by low-quality training samples, which can lead to inaccurate class centers and reduced model performance.

Purpose of the Study:

• To propose a novel system, High-Quality Class Center Learning (HQ2CL), for enhancing face recognition.
• To guide class centers towards high-quality samples, thereby preserving and improving model discriminability.
• To address the issue of inaccurate class centers caused by noisy or low-quality training data.

Main Methods:

• Introduction of a quality-aware scale and margin layer within the identification loss function.
• Development of a new high-quality center loss to refine the representation of identity centers.
• Implementation of the HQ2CL system without introducing additional computational overhead.

Main Results:

• Experimental evaluation on diverse face recognition benchmarks demonstrates the effectiveness of HQ2CL.
• The proposed system shows superior performance compared to existing state-of-the-art methods.
• HQ2CL successfully guides class centers toward high-quality samples, enhancing discriminability.

Conclusions:

• The HQ2CL system offers a significant improvement in face recognition by optimizing class center quality.
• The method effectively mitigates the negative impact of low-quality training samples on model performance.
• HQ2CL provides a computationally efficient and superior approach for robust face recognition.