U-Finger: Multi-Scale Dilated Convolutional Network for Fingerprint Image Denoising and Inpainting

Ramakrishna Prabhu, Xiaojing Yu, Zhangyang Wang, Ding Liu, Anxiao JiangarXiv 2018 · arXiv

Abstract

Fingerprint restoration is a vital task for biometric security and forensics, yet samples collected in real-world scenarios are often degraded by noise, occlusions, and complex backgrounds. We introduce U-Finger, a specialized U-shaped convolutional neural network designed for high-resolution fingerprint denoising and inpainting. By incorporating dilated convolutions and eliminating standard image padding, U-Finger maintains the integrity of fine ridge structures while suppressing artifacts. Our approach secured second place in the ECCV 2018 ChaLearn LAP competition, proving its efficacy in restoring severely distorted fingerprint data.

Introduction

Fingerprints are the gold standard for biometric identification, but "in the wild" samples are rarely pristine. They suffer from a cocktail of degradations: skin dryness, sensor noise, and physical scratches. Furthermore, fingerprints found at crime scenes are often overlaid on complex backgrounds like walls or stone, making them incredibly difficult to isolate. While deep learning has revolutionized natural image restoration, fingerprints present a unique challenge. Unlike standard photos, fingerprints are composed of fine-scale, high-contrast textures. Standard neural networks often "smudge" these textures, rendering the fingerprint useless for automated matching software.

Methodology

U-Finger employs a symmetric encoder-decoder structure across three scales, inspired by the U-Net architecture. The framework consists of two primary components:

  • Feature Encoding and Decoding: The network passes the image through three scales of cascaded feature modules. Each module consists of a convolutional layer followed by a residual block, spatial batch normalization, and ReLU activation. Skip connections help preserve spatial details that might otherwise be lost during downsampling.
  • Dilated Convolutions: Instead of standard convolutions, U-Finger uses a dilation factor of 3. This allows the network to have a larger receptive field—capturing the global flow of ridge patterns—without losing spatial resolution or increasing parameter count.
  • Zero-Padding Removal: Most CNNs use zero-padding to maintain image dimensions. However, the authors discovered that in fingerprint tasks, padding helps "leak" background noise into the final output. By removing all padding, the network is forced to focus strictly on internal structural data.
The U-Finger architecture (a) featuring multi-scale encoding/decoding and residual block configurations (b, c).

Experimental Results

The model was trained and evaluated using the ECCV 2018 ChaLearn LAP Inpainting Competition (Track 3) dataset, which contains realistic artificial fingerprints degraded with various distortions and overlaid on diverse backgrounds.

Quantitative Performance

U-Finger was a top performer in the competition, securing the overall No. 2 ranking. On the hold-out testing set, it achieved:

  • MSE: 0.0231 (Rank 2)
  • PSNR: 16.9688 dB (Rank 2)
  • SSIM: 0.8093 (Rank 3)

Qualitative Analysis

Qualitatively, the model showed a remarkable ability to suppress complex background textures (like stone grain) while keeping the fingerprint ridges sharp and continuous. The model performs "blind inpainting," meaning it learns to recognize and reconstruct broken ridge patterns based on the surrounding global flow without being explicitly told which parts are missing.

Qualitative results showing (a) a fingerprint on a complex stone background, (b) the U-Finger restoration, and (c) the ground truth.

Conclusion

U-Finger demonstrates that tailoring architectural choices—specifically dilation and the removal of padding—to the unique structural properties of fingerprints can significantly improve restoration quality. While the model excels at structural restoration, future work may explore alternative loss functions, such as perceptual or adversarial losses, to further sharpen the visual results and validate the model on a wider variety of actual latent fingerprints from diverse surfaces.

Summary updated

BibTeX
@misc{prabhu2018u,
  author = {Ramakrishna Prabhu and Xiaojing Yu and Zhangyang Wang and Ding Liu and Anxiao Jiang},
  title = {{U-Finger: Multi-Scale Dilated Convolutional Network for Fingerprint Image Denoising and Inpainting}},
  year = {2018},
  eprint = {1807.10993},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.10993},
}
Figure