Efficient video face recognition based on frame selection and quality assessment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Video face identification accuracy is limited by suboptimal frame selection.
• Traditional methods for frame quality assessment (e.g., brightness, contrast) are insufficient.
• Existing advanced models like FaceQNet are effective but lack accessible training data.

Purpose of the Study:

• To improve the performance and accuracy of video face identification.
• To develop a novel method for selecting the best video frames for identification.
• To introduce an efficient frame quality assessment technique using deep learning.

Main Methods:

• Utilizing a lightweight convolutional neural network for frame quality estimation.
• Applying knowledge distillation from the FaceQNet model to train the lightweight network.
• Selecting the K-best frames and generating a single average descriptor for classification.
• Comparing the proposed algorithm against traditional frame-by-frame feature extraction and clustering methods.

Main Results:

• The proposed deep learning approach for frame quality assessment outperforms traditional methods.
• Knowledge distillation enables effective training of a lightweight model without large public datasets.
• The K-best frame selection strategy combined with average descriptors improves identification accuracy.
• The algorithm demonstrates superior performance compared to existing video face identification techniques.

Conclusions:

• Deep learning-based frame quality assessment is a viable strategy for enhancing video face identification.
• Knowledge distillation offers a practical solution for adapting complex models to specific tasks with limited data.
• The developed method provides a more robust and accurate approach to video face identification.