Facial identity recognition using StyleGAN3 inversion and improved tiny YOLOv7 model.

Area of Science:

• Computer Vision and Artificial Intelligence
• Biometrics and Security

Background:

• Facial identity recognition is complex due to attribute manipulation, often exploited in criminal activities.
• Existing methods struggle with few-shot learning and attribute variations.

Purpose of the Study:

• To propose a one-step deep learning solution for facial attribute manipulation detection and facial identity recognition.
• To enhance recognition accuracy in both few-shot and traditional learning scenarios.

Main Methods:

• Creation of the Facial Attribute Manipulation Detection (FAM) Dataset with 11,560 images across 20 identities and 38 attributes.
• Development of the FIR-Tiny YOLOv7 model, integrating Spatial Transformer Block (STB) and Squeeze-Excite Spatial Pyramid Pooling (SE-SPP) into Tiny YOLOv7.
• Utilized StyleGAN3 inversion for attribute generation and YOLO format for annotation.

Main Results:

• The FIR-Tiny YOLOv7 model achieved significant improvements in mean Average Precision (mAP): 10.0% (one-shot), 30.4% (three-shot), and 15.3% (five-shot).
• A marginal improvement of 0.1% mAP was observed in the traditional 70%-30% split scenario compared to baseline Tiny YOLOv7.
• Demonstrated superior performance in recognizing facial identities despite attribute manipulation.

Conclusions:

• The proposed FIR-Tiny YOLOv7 model offers a promising approach for robust facial identity recognition.
• Effective in handling scenarios with varying facial attribute manipulations and limited data.
• The FAM dataset provides a valuable resource for research in facial attribute manipulation detection.