Scholarlink

<p>Identifying deceased individuals in forensic and humanitarian contexts, particularly those affected by torture and trauma, remains a significant challenge due to severe facial damage, decomposition, and substantial image disparities between antemortem (AM) and postmortem (PM) records. To address this issue, a novel automated framework for robust postmortem face recognition is proposed in this paper. The proposed methodology integrates a dual-method face detection strategy with automatic orientation correction to achieve a 95% detection rate on challenging PM images, drastically reducing manual intervention. For recognition, we leverage the quality-adaptive capabilities of AdaFace-ViT to generate discriminative embeddings. Extensive evaluation on a difficult dataset of AM and PM facial images shows the outstanding performance of the proposed framework. AdaFace-ViT outperformed FaceNet by 14.2% in top-5 accuracy and was 10<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\times\)</EquationSource></InlineEquation> more accurate with 3<InlineEquation ID="IEq2"><EquationSource Format="TEX">\(\times\)</EquationSource></InlineEquation> lower EER than traditional CNNs. The obtained results demonstrate the robustness of the proposed framework to extreme facial deterioration and its potential to enable scalable, reliable humanitarian identification systems.</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

Quality-adaptive forensic face recognition using a dual-detector pipeline with AdaFace-ViT for postmortem identification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • A. H. Abdul Hafez,
  • Ahmed El Jouma

摘要

Identifying deceased individuals in forensic and humanitarian contexts, particularly those affected by torture and trauma, remains a significant challenge due to severe facial damage, decomposition, and substantial image disparities between antemortem (AM) and postmortem (PM) records. To address this issue, a novel automated framework for robust postmortem face recognition is proposed in this paper. The proposed methodology integrates a dual-method face detection strategy with automatic orientation correction to achieve a 95% detection rate on challenging PM images, drastically reducing manual intervention. For recognition, we leverage the quality-adaptive capabilities of AdaFace-ViT to generate discriminative embeddings. Extensive evaluation on a difficult dataset of AM and PM facial images shows the outstanding performance of the proposed framework. AdaFace-ViT outperformed FaceNet by 14.2% in top-5 accuracy and was 10\(\times\) more accurate with 3\(\times\) lower EER than traditional CNNs. The obtained results demonstrate the robustness of the proposed framework to extreme facial deterioration and its potential to enable scalable, reliable humanitarian identification systems.