A Deep Generative Recognition Framework for Low-Quality and Partial Fingerprints
Jainy Jacob M, D. Shanmugapriya · 2025
Fingerprint recognition systems are key elements in biometric authentication systems and forensic investigations; however, their performance is severely degraded in the presence of low-quality or partial fingerprints. These degradations are caused by sensor limitations, environmental noise, smudging, or incomplete impressions, which then produce instances of incorrect identification or verification. To overcome these limitations, develop an innovative Deep Generative Recognition Network (DGR-Net), a unified approach to generative reconstruction and discriminative recognition. DGR-Net uses a deep encoder-decoder framework to infer and regenerate potentially missing or damaged ridge patterns in partial fingerprints. The output of the DGR-Net model is a reconstructed image that is then supplied to the DGR-Net generative model convolution feature extractor to ensure important features of finger prints, such as minutiae points or ridge orientation, are maintained. The proposed architecture is trained end-to-end using combinations of synthetic and real-world low-quality fingerprint datasets using a composite loss function that consists of reconstruction loss, identity matching loss, and structural consistency. In experimentation shows that DGR-Net outperforms competing paradigms using U-Net variants, VAE and GAN-based models, with respect to PSNR, SSIM, and Rank-1 recognition accuracy with 30 dB PSNR, 0.88 SSIM, and 48% Rank-1 recognition accuracy. In addition, our model achieves reasonably high minutiae precision and recall, with improved NFIQ scores that show substantial improvement in biometric utility. We also identify DGR-Net as a strong candidate for deployment in difficult fingerprint recognition environments, representing a viable contribution in security-critical contexts and forensic applications.