Cross-Batch Hard Example Mining With Pseudo Large Batch for ID vs. Spot Face Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Identity verification systems commonly use ID vs. Spot (IvS) face recognition, matching ID photos to live images.
• IvS face recognition presents unique challenges due to massive class numbers and only two images per class, straining GPU memory and feature extraction.
• Existing methods struggle with the scale and data scarcity inherent in IvS face recognition datasets.

Purpose of the Study:

• To develop an effective and memory-efficient model for ID vs. Spot (IvS) face recognition.
• To enhance feature extraction capabilities for bisample identity data in large-scale face recognition.
• To address the computational challenges posed by massive datasets in identity verification.

Main Methods:

• A two-stage training approach: initial training on general face recognition datasets, followed by metric learning on IvS data.
• Introduction of Cross-Batch Hard Example Mining (CB-HEM) to expand hard triplet selection across multiple mini-batches.
• Proposal of Pseudo Large Batch (PLB) to virtually increase batch size, optimizing GPU memory usage.
• Simultaneous application of CB-HEM and PLB to significantly enlarge the sample selection space for training.

Main Results:

• The proposed CB-HEM and PLB methods effectively enhance the discriminative capability of face recognition models.
• These algorithms enable the selection of very hard sample pairs, including hard negatives, crucial for robust identity verification.
• Extensive evaluations on multiple IvS benchmarks confirm the superior performance of the developed method compared to existing approaches.

Conclusions:

• The combination of CB-HEM and PLB offers a significant advancement in ID vs. Spot face recognition.
• The proposed methods provide a scalable and efficient solution for training models on large-scale, bisample identity datasets.
• This research contributes to more accurate and reliable identity verification systems for real-world applications like ePassport gates.