Seeing through disguise: Getting to know you with a deep convolutional neural network

Area of Science:

• Cognitive Science
• Computer Vision
• Artificial Intelligence

Background:

• Disguise presents challenges for face recognition, impacting both identity consistency (evasion) and differentiation (impersonation).
• Personal familiarity enhances human ability to recognize faces despite disguise.
• Deep convolutional neural networks (DCNNs) offer a computational framework for modeling high-level visual learning in face identification.

Purpose of the Study:

• To propose and test a computational model of facial familiarity using DCNNs.
• To investigate how DCNNs process disguise and familiarity.
• To explore mechanisms for grouping same-identity images and separating different identities.

Main Methods:

• Trained a DCNN for face identification.
• Assessed DCNN performance on disguised faces (unfamiliar condition).
• Simulated familiarity by averaging identity representations and employed contrast learning techniques.

Main Results:

• DCNN performance was degraded by disguise, mirroring human responses.
• Simulated familiarity improved recognition for evasion disguise but hindered differentiation of similar identities.
• Contrast learning enhanced DCNN performance for both evasion and impersonation disguises.

Conclusions:

• DCNNs can model the effects of disguise on face recognition.
• Simulated familiarity and contrast learning are crucial for robust face recognition models.
• High-level visual representations in DCNNs support simultaneous clustering and separation of identities, crucial for familiar face recognition.