Scholarlink

<p>In this work, we present a pioneering face blending data augmentation technique that enhances facial recognition performance, particularly when dealing with limited datasets. By applying a blending method to intra-class facial features, intermediate faces that retain the intrinsic characteristics of their original class are generated. We illustrate the efficacy of this approach through the integration of the MTCNN model and Warping-based interpolation. The generated faces are evaluated qualitatively, followed by a rigorous quantitative analysis using metrics such as Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Inception Score (IS), and VGG-based Similarity. Extensive experimental results demonstrate that our Data augmentation approach significatively enhance the classification performanceof the InceptionV3 model, surpassing several augmentation methods in terms of accuracy, precision, recall, and F1 score. This approach offers significant potential for enhancing the robustness and accuracy of facial classification systems, making it particularly effective for real-world applications constrained by limited data availability.</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

Revolutionizing facial recognition: Boosting performance on limited data with InceptionV3-based face blending

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Emna Ghorbel,
  • Ghada Maddouri,
  • Faouzi Ghorbel

摘要

In this work, we present a pioneering face blending data augmentation technique that enhances facial recognition performance, particularly when dealing with limited datasets. By applying a blending method to intra-class facial features, intermediate faces that retain the intrinsic characteristics of their original class are generated. We illustrate the efficacy of this approach through the integration of the MTCNN model and Warping-based interpolation. The generated faces are evaluated qualitatively, followed by a rigorous quantitative analysis using metrics such as Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Inception Score (IS), and VGG-based Similarity. Extensive experimental results demonstrate that our Data augmentation approach significatively enhance the classification performanceof the InceptionV3 model, surpassing several augmentation methods in terms of accuracy, precision, recall, and F1 score. This approach offers significant potential for enhancing the robustness and accuracy of facial classification systems, making it particularly effective for real-world applications constrained by limited data availability.