Scholarlink

<p> <?tk 1?>This paper presents a novel end-to-end emotion-preserving face de-identification approach based on Generative Adversarial Networks (GANs), specifically utilizing the StyleGAN architecture. The proposed method generates natural-looking de-identified images by creating a synthetic face dataset and leveraging the DeepFace model for gender classification and representative image selection. To enhance emotion preservation, an improved SimSwap framework is introduced, incorporating a novel loss function designed to maintain emotional expressions. The DeepFace model is further utilized to classify and recognize emotional expressions in both original and de-identified images. Emotion preservation during face swapping is explicitly enforced by minimizing attribute preservation loss. A comprehensive ablation study demonstrates the effectiveness of the proposed GAN components. Experimental results show that the method outperforms recent face de-identification techniques in both accuracy and emotion preservation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Emotion-aware face de-identification with generative adversarial networks

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • Md Shopon,
  • Marina L. Gavrilova

摘要

This paper presents a novel end-to-end emotion-preserving face de-identification approach based on Generative Adversarial Networks (GANs), specifically utilizing the StyleGAN architecture. The proposed method generates natural-looking de-identified images by creating a synthetic face dataset and leveraging the DeepFace model for gender classification and representative image selection. To enhance emotion preservation, an improved SimSwap framework is introduced, incorporating a novel loss function designed to maintain emotional expressions. The DeepFace model is further utilized to classify and recognize emotional expressions in both original and de-identified images. Emotion preservation during face swapping is explicitly enforced by minimizing attribute preservation loss. A comprehensive ablation study demonstrates the effectiveness of the proposed GAN components. Experimental results show that the method outperforms recent face de-identification techniques in both accuracy and emotion preservation.