Low-resolution Face Recognition in the Wild via Selective Knowledge Distillation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deploying face recognition in real-world scenarios requires identifying low-resolution faces efficiently.
• Existing complex models are computationally expensive, limiting their practical application.

Purpose of the Study:

• To develop a computationally efficient method for recognizing low-resolution faces.
• To compress complex face recognition models with minimal performance degradation.

Main Methods:

• A two-stream convolutional neural network (CNN) architecture with a teacher (complex) and student (simple) stream.
• Selective knowledge distillation using sparse graph optimization to transfer informative features from teacher to student.
• Regularizing the student stream by simultaneously performing feature regression and low-resolution face classification.

Main Results:

• The proposed method enables impressive performance in low-resolution face recognition.
• The student stream achieves a low memory footprint (0.15MB) and high processing speeds (418 FPS on CPU, 433 FPS on GPU).

Conclusions:

• Selective knowledge distillation is an effective strategy for creating efficient and accurate low-resolution face recognition models.
• The developed approach offers a feasible solution for deploying face recognition in resource-constrained environments.