Scholarlink

<p>Forensic biometrics, latent fingerprint enhancement is still viewed as a central challenge, as crime scene impressions are notorious for poor quality, such that impressions are noisy, smudged, and/or missing many ridge structures. However, existing deep learning methods good at visual fidelity do not retain global fingerprint topology and even hallucinate spurious ridge patterns in the degraded regions. This paper proposes a novel Dirichlet-Augmented Conditional Generative Adversarial Network (DA-cGAN) that includes semantic fingerprint structure via unsupervised Dirichlet-Multinomial clustering of local ridge features such as orientation, frequency, thickness and confidence. These semantic clusters guide a conditional GAN to apply region-aware enhancement, which is additionally reinforced by a new semantic cluster consistency loss that enforces coherent enhancement within structurally similar regions. Experimental results are provided using three benchmark datasets—IIITD MOLF (multi-sensor optical and latent fingerprint), IIITD Latent and NIST SD27, showing that the performance of DA–cGAN is superior. MOLF (DB1) and IIITD Latent are ranked at positions 74.92% and 100%. Furthermore, according to image quality metrics, DA-cGAN produces the highest quality images: 28.53 dB PSNR and 0.911 SSIM on MOLF and 29.32 dB PSNR (peak signal-to-noise ratio) and 0.925 SSIM (structural similarity index measure) on Latent, respectively.</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

DA-cGAN: Dirichlet-Augmented Conditional GAN for structure-aware latent fingerprint enhancement

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Nandita Manchanda,
  • Sanjay Singla

摘要

Forensic biometrics, latent fingerprint enhancement is still viewed as a central challenge, as crime scene impressions are notorious for poor quality, such that impressions are noisy, smudged, and/or missing many ridge structures. However, existing deep learning methods good at visual fidelity do not retain global fingerprint topology and even hallucinate spurious ridge patterns in the degraded regions. This paper proposes a novel Dirichlet-Augmented Conditional Generative Adversarial Network (DA-cGAN) that includes semantic fingerprint structure via unsupervised Dirichlet-Multinomial clustering of local ridge features such as orientation, frequency, thickness and confidence. These semantic clusters guide a conditional GAN to apply region-aware enhancement, which is additionally reinforced by a new semantic cluster consistency loss that enforces coherent enhancement within structurally similar regions. Experimental results are provided using three benchmark datasets—IIITD MOLF (multi-sensor optical and latent fingerprint), IIITD Latent and NIST SD27, showing that the performance of DA–cGAN is superior. MOLF (DB1) and IIITD Latent are ranked at positions 74.92% and 100%. Furthermore, according to image quality metrics, DA-cGAN produces the highest quality images: 28.53 dB PSNR and 0.911 SSIM on MOLF and 29.32 dB PSNR (peak signal-to-noise ratio) and 0.925 SSIM (structural similarity index measure) on Latent, respectively.