[Paper Review] Deep End-to-end Fingerprint Denoising and Inpainting

Youness Mansar|arXiv (Cornell University)|Jul 31, 2018

Digital Media Forensic Detection4 references4 citations

TL;DR

This paper presents an end-to-end deep learning approach using a U-Net-like convolutional neural network to denoise and inpaint synthetic fingerprint images, achieving state-of-the-art performance on the Chalearn LAP In-painting Competition Track 3. The method leverages data augmentation, adaptive learning rate scheduling, and skip connections, outperforming all competitors on MAE, PSNR, and SSIM metrics.

ABSTRACT

This work describes our winning solution for the Chalearn LAP In-painting Competition Track 3 - Fingerprint Denoising and In-painting. The objective of this competition is to reduce noise, remove the background pattern and replace missing parts of fingerprint images in order to simplify the verification made by humans or third-party software. In this paper, we use a U-Net like CNN model that performs all those steps end-to-end after being trained on the competition data in a fully supervised way. This architecture and training procedure achieved the best results on all three metrics of the competition.

Motivation & Objective

• To develop a robust, end-to-end deep learning model that simultaneously denoises and inpaints degraded fingerprint images.
• To improve fingerprint verification accuracy by reducing false acceptance and rejection rates through image pre-processing.
• To address challenges posed by synthetic distortions such as blur, occlusion, noise, and background patterns in fingerprint images.
• To evaluate the effectiveness of architectural components like skip connections and data augmentation in a fully supervised learning setup.
• To demonstrate generalization potential of a CNN-based approach beyond image segmentation tasks in fingerprint restoration.

Proposed method

• A U-Net-like encoder-decoder convolutional neural network is used, with skip connections to preserve spatial details and mitigate vanishing gradient issues.
• Input images are normalized to [0,1] and resized to dimensions divisible by 16 (2^4) to accommodate max-pooling layers.
• Data augmentation is applied during training only and includes random flips, shears, translations, zooms, contrast/saturation changes, and rotations.
• The model is trained using the Adam optimizer with an initial learning rate of 1e-4, reduced by a factor of 0.5 when validation loss plateaus for 3 epochs.
• Output images are min-max scaled to [0,255] and resized back to the original input dimensions before submission.
• The loss function is based on Mean Absolute Error (MAE), which serves as a strong proxy for PSNR and SSIM during training.

Experimental results

Research questions

• RQ1 Can a single, end-to-end deep learning model effectively handle both denoising and inpainting in fingerprint images without task-specific pre-processing?
• RQ2 How do architectural components like skip connections and data augmentation influence performance on synthetic fingerprint restoration?
• RQ3 To what extent does adaptive learning rate scheduling improve convergence and final metric performance?
• RQ4 Does training with only synthetic data generalize to real-world fingerprint image restoration?
• RQ5 Can MAE loss alone yield strong performance on PSNR and SSIM, indicating its suitability as a proxy loss for fingerprint image quality?

Key findings

• The proposed model achieved the best results on all three competition metrics: MAE of 0.0189, PSNR of 17.6968, and SSIM of 0.8427 on the test set.
• The ablation study showed that reducing the learning rate improved performance the most, while data augmentation had minimal impact.
• Removing skip connections degraded performance drastically, with MAE rising to 0.7282 and SSIM dropping to 0.0001.
• Using a pre-trained VGG encoder improved results slightly, reducing MAE to 0.0225 and increasing SSIM to 0.8142.
• Despite using only MAE in the loss function, the model generalized well to PSNR and SSIM, indicating MAE acts as an effective proxy loss.
• The model outperformed other approaches, including those using dilated convolutions (rgsl888) and M-Net architectures (sukeshadigav), across all metrics.