Deep learning features in facial identification and the likelihood ratio bound.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Score-based likelihood ratio (SLR) is a key method in facial comparison.
• Deep learning (DL) features are increasingly used in facial recognition.
• The theoretical underpinnings of DL features in SLR methods require further investigation.

Purpose of the Study:

• To investigate deep learning facial features and their Score-based likelihood ratio (SLR) levels.
• To propose a novel interpretation of deep learning features as class characteristics.
• To evaluate the effectiveness of DL features in facial identification using SLR.

Main Methods:

• Utilized a large-scale dataset for experimental analysis.
• Calculated match scores from deep learning-based facial recognition systems.
• Analyzed the Score-based likelihood ratio (SLR) values derived from these features.

Main Results:

• Presented evidence that log SLR values for deep learning features can reach 8 in specific datasets.
• Demonstrated that deep learning features can be interpreted as class characteristics.
• Quantified the SLR levels associated with deep learning facial features.

Conclusions:

• The Score-based likelihood ratio (SLR) of deep learning features is a valuable metric for facial identification.
• This method is particularly effective for identifying individuals from CCTV camera images.
• Deep learning features, when viewed as class characteristics, enhance facial recognition accuracy within SLR frameworks.