LittleFaceNet: A Small-Sized Face Recognition Method Based on RetinaFace and AdaFace

Area of Science:

• Computer Vision
• Biometrics
• Artificial Intelligence

Background:

• Traditional face recognition struggles with low-resolution and occluded images common in surveillance.
• Existing target detection algorithms require extensive data, which is scarce for low-resolution faces.
• University laboratory surveillance faces unique challenges like moving occlusions and poor image quality.

Purpose of the Study:

• To develop a robust face recognition framework for low-resolution and occluded faces in surveillance.
• To improve the accuracy of face detection and recognition in challenging laboratory environments.
• To address the limitations of existing algorithms in handling scarce low-resolution face datasets.

Main Methods:

• Reconstruction of Retinaface-Resnet for enhanced face detection and localization.
• Integration of Quality-Adaptive Margin (adaface) for low-resolution face recognition using feature norm approximation.
• Introduction of Spatial Depth-wise Separable Convolutions to improve detection of small and angled faces.
• Implementation of a multi-object tracking algorithm to handle moving occlusions.

Main Results:

• Achieved 96.12% accuracy on the challenging WiderFace dataset.
• Demonstrated 84.36% recognition accuracy in practical laboratory surveillance applications.
• Significant improvements in handling low-resolution, occluded, and extreme-angle face scenarios.

Conclusions:

• The proposed Retinaface-Resnet and adaface framework effectively addresses key challenges in low-resolution face recognition.
• The integration of spatial convolutions and multi-object tracking enhances robustness in real-world surveillance.
• This framework offers a viable solution for accurate person identification in constrained environments like university labs.