Analysis of Real-Time Face-Verification Methods for Surveillance Applications

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Deep learning face recognition uses complex models, limiting deployment on resource-constrained devices.
• Lightweight methods offer real-time performance but struggle with unconstrained surveillance conditions like low resolution and face rotation.

Purpose of the Study:

• To compare three state-of-the-art (SOTA) real-time face verification methods for surveillance applications.
• To evaluate method performance against challenges such as face rotation and low-resolution images.

Main Methods:

• Created a 3000-image evaluation subset with varied face rotation and resolution levels from existing datasets.
• Methodically evaluated MobileFaceNet, EfficientNet-B0, and GhostNet on the custom subset and conventional datasets (Cross-Pose LFW, QMUL-SurvFace).

Main Results:

• EfficientNet-B0 demonstrated capability in handling both low-resolution and face rotation issues common in surveillance.
• MobileFaceNet showed superior performance in extreme face rotation scenarios, exceeding 80 degrees.

Conclusions:

• EfficientNet-B0 is a strong candidate for general surveillance face verification.
• MobileFaceNet is recommended for applications requiring robust performance under severe face rotation conditions.