[Paper Review] Latent fingerprint minutia extraction using fully convolutional network

Yao Tang, Fei Gao+1 more|arXiv (Cornell University)|Sep 30, 2016

Biometric Identification and Security15 references3 citations

TL;DR

This paper proposes a fully convolutional network (FCN)-based method for end-to-end latent fingerprint minutiae extraction, bypassing traditional preprocessing steps. By generating pixel-level minutiae score maps and refining proposals with a shared CNN for classification and orientation regression, the approach achieves 53% recall and 53% precision on NIST SD27, outperforming prior methods while preserving ridge structure through feature visualization.

ABSTRACT

Minutiae play a major role in fingerprint identification. Extracting reliable minutiae is difficult for latent fingerprints which are usually of poor quality. As the limitation of traditional handcrafted features, a fully convolutional network (FCN) is utilized to learn features directly from data to overcome complex background noises. Raw fingerprints are mapped to a correspondingly-sized minutia-score map with a fixed stride. And thus a large number of minutiae will be extracted through a given threshold. Then small regions centering at these minutia points are entered into a convolutional neural network (CNN) to reclassify these minutiae and calculate their orientations. The CNN shares convolutional layers with the fully convolutional network to speed up. 0.45 second is used on average to detect one fingerprint on a GPU. On the NIST SD27 database, we achieve 53\% recall rate and 53\% precise rate that outperform many other algorithms. Our trained model is also visualized to show that we have successfully extracted features preserving ridge information of a latent fingerprint.

Motivation & Objective

• To address the challenge of reliable minutiae extraction in low-quality latent fingerprints, which often fail under traditional handcrafted feature methods.
• To eliminate reliance on manual segmentation, enhancement, or pre-processing by learning features directly from raw fingerprint images.
• To improve detection accuracy by modeling minutiae as point detection with joint regression of location and orientation.
• To enable end-to-end learning of robust minutiae descriptors that preserve ridge structure and reduce noise sensitivity.

Proposed method

• A fully convolutional network (FCN) maps raw fingerprints to a minutiae-score map with a fixed stride of 16, generating pixel-level proposals based on learned scores.
• Proposals are filtered using a threshold to identify candidate minutiae locations, with location regression used as a multi-task loss to refine spatial accuracy.
• A shared convolutional neural network (CNN) classifies proposals as true minutiae and regresses their orientation using a multi-task loss layer.
• The FCN and CNN share early convolutional layers to accelerate inference and ensure feature consistency across stages.
• Feature visualization via activation reconstruction confirms that the network learns to preserve ridge and valley patterns in latent fingerprints.
• The model is trained end-to-end on raw fingerprint images using a combination of classification, location, and orientation regression losses.

Experimental results

Research questions

• RQ1 Can a fully convolutional network effectively generate reliable minutiae proposals directly from raw, low-quality latent fingerprints without segmentation or enhancement?
• RQ2 Does end-to-end learning of minutiae descriptors through deep networks outperform handcrafted feature methods in noisy, complex backgrounds?
• RQ3 To what extent do joint regression of location and orientation improve minutiae detection accuracy compared to classification alone?
• RQ4 Can the learned features preserve ridge structure and distinguish minutiae from false positives in challenging latent fingerprint conditions?

Key findings

• The proposed method achieves a 53% recall and 53% precision rate on the NIST SD27 database, outperforming multiple existing algorithms.
• The model processes one fingerprint in an average of 0.45 seconds on a GPU, demonstrating high efficiency.
• Ablation studies show that removing location regression reduces recall and precision to 50.0% and 49.7%, respectively, indicating its importance for localization accuracy.
• Removing orientation regression increases recall and precision to 55.0% and 55.4%, suggesting orientation filtering helps eliminate false positives.
• Visualization results confirm that the network learns to extract and preserve ridge and valley structures, with significant improvement after fine-tuning on latent fingerprints.
• The method achieves better identification performance than both mindtct and Gabor&CNN-based approaches when evaluated on 129 latent fingerprints from NIST SD27.