Frankenstein: Learning Deep Face Representations Using Small Data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) excel at face recognition in uncontrolled environments.
• Training these CNNs requires massive labeled image datasets, often millions of images.
• Large datasets are unavailable for specific applications like near-infrared (NIR) face recognition.

Purpose of the Study:

• To develop a method for generating large synthetic training datasets for face recognition.
• To address the data scarcity issue in specialized domains like NIR face recognition.
• To enable effective model training with significantly reduced data requirements.

Main Methods:

• Proposing a novel method to generate synthetic face images by compositing real images.
• Utilizing composited images to create large-scale training datasets.
• Training deep convolutional neural networks on these synthetic datasets.

Main Results:

• Models trained on as few as 10,000 synthetic images achieve performance comparable to models trained on 500,000 real images.
• The proposed method effectively overcomes the limitations of data scarcity in NIR face recognition.
• State-of-the-art results were achieved on the CASIA NIR-VIS2.0 heterogeneous face recognition dataset.

Conclusions:

• Synthetic data generation via image compositing is a viable solution for training deep CNNs in data-scarce domains.
• This approach significantly reduces the need for extensive data collection efforts.
• The method holds promise for advancing face recognition technologies, particularly in challenging conditions like NIR spectrum.