Data Augmentation for Facial Recognition with Diffusion Model

CVPR 2024 Workshop SyntaGen Submission3 Authors

Published: 07 Apr 2024, Last Modified: 12 Apr 2024SyntaGen 2024EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Diffusion Model, Data Augmentation, Facial Recognition
TL;DR: The paper proposes a new image generation pipeline based on ControlNet for facial data synthesis
Abstract: In recent years, facial recognition technology has made significant progress. However, it also faces challenges in common scenarios of daily life. For example, facial accessories such as masks, glasses, and hats have a negative impact on recognition accuracy. This paper introduces a facial data synthesis pipeline based on the diffusion model, which combines the text-to-image generation method with Mask-ControlNet. The pipeline can generate various common facial occlusions, achieving diverse and high-fidelity facial image generation. By comparing the performance of different models trained with synthetic and real images, extensive experimental results confirm the effectiveness of this method in enhancing the robustness of facial recognition.
Submission Number: 3
Loading