Clean-View Distillation for Rectifying Distorted Fingerprints Without Skeletons

When a finger presses on a sensor, the skin stretches, twists and compresses unevenly. The ridge pattern in the image is then warped relative to the same finger's enrolled print, and a matcher that expects rigid alignment rejects true matches. The best current fix for a single print estimates a dense distortion field, meaning a per-pixel map of how far each point was pushed, and then warps the print back. The method in Regression of Dense Distortion Field from a Single Fingerprint Image does this well on clean prints. It depends on a preprocessing step, though: it first thins the ridges to one-pixel-wide lines, a step called skeletonization. The authors report that this step breaks on smudged, low-contrast or noisy prints, and those are exactly the crime-scene latent prints where rectification is most needed.

We propose to train a model that reads raw grayscale pixels with a teacher that reads clean, skeletonized images. Take a clean print and corrupt only its appearance with smudge, noise, background clutter and partial cropping, leaving the geometry alone. The distortion field stays the same, because the skin was stretched the same way. The skeleton-based teacher sees the clean copy, where its preprocessing works, and produces a per-pixel distribution over displacements. A single-stream student sees the corrupted copy and is trained to reproduce that distribution. This is knowledge distillation, training a small or differently-equipped model to imitate a stronger one. Here the purpose is to move a capability from easy inputs to hard ones, more than to compress the model. Because the teacher supplies the targets, training can use any collection of clean prints, including ones that have no traced ground-truth displacement.

Core Claim

A distortion field is a geometric quantity, and it is unchanged by photometric degradation. A teacher's field estimate on a clean image is therefore a valid training target for every degraded rendering of that image. Distilling soft, interval-binned displacement distributions across this clean-to-degraded gap should produce a skeleton-free student with three properties. It should roughly match the teacher on clean prints. It should beat the teacher on degraded and latent prints. It should also beat a student trained only on the available ground-truth fields with the same augmentations. Prior distillation work does not test the last property, and it is what would make the idea more than augmentation.

The method draws on two prior lines of work. The first is dense-output distillation. Structured Knowledge Distillation for Dense Prediction and Channel-wise Knowledge Distillation for Dense Prediction* transfer pixel-wise distributions and spatial affinities for segmentation and detection. They have never been applied to displacement fields, and the teacher and student always see the same input. The second is the interval formulation of Phase-Aggregated Dual-Branch Network for Efficient Fingerprint Dense Registration, which casts each pixel's displacement as a 25-bin classification. That paper trains on hard bin labels. Its formulation also means a teacher can emit soft distributions whose spread encodes ambiguity in the deformation, which is the soft-label idea of Transferring knowledge from a RNN to a DNN applied to geometry. The closest biometric distillation work, Knowledge Distillation from Multi-Modality to Single-Modality for Person Verification, transfers multimodal identity embeddings and does not touch deformation.

Research Program

The teacher is a reimplementation of the single-image regressor from Regression of Dense Distortion Field from a Single Fingerprint Image. It keeps the texture and orientation branches, the skeleton input and the spatial pyramid attention. Its output head is replaced with a per-pixel, per-axis interval classifier in the style of PDRNet. It is trained on the video-traced ground-truth fields of the Tsinghua distorted database (TDF/TDF_V2), as in Dense Registration and Mosaicking of Fingerprints by Training an End-to-End Network. We first confirm that the reimplementation reproduces the published minutiae reprojection error, which measures how far rectified minutiae land from their true positions.

The student is a single-stream encoder-decoder that takes raw grayscale pixels, with no skeleton and no orientation preprocessing. For each clean training print $x$, we sample a photometric corruption $c$, such as blur, smudge, ridge breaks, structured background, contrast loss or partial masking. The student is trained on

$$ \mathcal{L} = \sum_{p} w_p \,\mathrm{KL}\big(q_T(\cdot \mid x, p) \,\|\, q_S(\cdot \mid c(x), p)\big) + \lambda\, \mathcal{L}_{\mathrm{GT}}, $$

where $q_T$ and $q_S$ are the teacher's and student's distributions over displacement bins at pixel $p$, and $w_p$ is a confidence weight that shrinks toward zero where the teacher's distribution has high entropy or the clean skeleton is poor. $\mathcal{L}_{\mathrm{GT}}$ is the ordinary bin-classification loss, applied only to prints that have traced ground truth, and $\lambda$ sets its weight. The unlabeled pool consists of clean prints from training partitions of FVC2004 and untraced TDF frames, kept disjoint from all evaluation subjects. All corruptions are photometric or masking operations, so the label-invariance assumption holds by construction. Whether the synthetic corruptions resemble real latents is an empirical question, and the evaluation on real latent prints is designed to answer it.

The analysis has three parts. We stratify errors by clean-skeleton quality to see where the student's advantage concentrates. We compare soft against hard teacher targets. We test whether the student's predicted entropy flags pixels where rectification is unreliable, which would let a matcher down-weight them.

Baselines

The naive reference point is no rectification: a minutiae matcher applied to the raw print. Every rectification method has to improve on this, and on latent prints it shows how much distortion actually costs.

The strongest existing method is the published skeleton-based single-image regressor, run as published including its skeletonization step, which is also our teacher. On TDF_V2 that paper reports a minutiae mean reprojection error reduced from 12.17 to 9.60. On the Hisign Latent Database it reports Rank-1 identification improving from 91.2% to 93.2% after rectification (Regression of Dense Distortion Field from a Single Fingerprint Image). For NIST SD27, no reported Rank-1 number for this method was found on Lacuna, and PDRNet's SD27 gain is reported only qualitatively. We rerun the teacher ourselves with the same matcher and the same evaluation splits as the student, so that matcher differences cannot explain a gap. Beating it means being no worse than about 0.3 pixels of reprojection error on clean TDF_V2 and clearly better on degraded TDF_V2 and on HLD and SD27 Rank-1 under the shared matcher.

The key ablation is a ground-truth-only student. It uses the identical architecture, the identical corruption pipeline and the same training compute and tuning budget, and it is trained only on traced TDF fields with no teacher. A second ablation gives the teacher the corrupted image instead of the clean one, which is ordinary self-distillation; this isolates the clean-view gap. A third replaces soft bins with the teacher's argmax bin, which isolates the soft targets. The pairwise registration networks, PDRNet and the end-to-end Siamese model, are reported as context on FVC2004 DB1_A, where PDRNet reports a normalized cross-correlation of 0.65 between registered images (Phase-Aggregated Dual-Branch Network for Efficient Fingerprint Dense Registration). They need a reference print and solve a different problem, so they are not the main comparison.

Why not just…?

Why not just train the raw-image student on ground-truth fields with heavy augmentation? This is the most serious threat, and the ground-truth-only ablation tests it directly. The teacher can add value in only two ways: more training prints through pseudo-labels on untraced data, and soft distributions that carry ambiguity. If the ablation matches the distilled student when both use only traced data, then the soft targets contribute nothing. We would then report the gain, if any, as coming entirely from unlabeled scale, and test it by varying the size of the unlabeled pool. Distiller found in language tasks that augmentation alone trails representation-level distillation, but that is weak evidence for geometry, so we rely on the experiment rather than the analogy.

Why not just put a fingerprint enhancement or denoising front-end before the existing teacher? Enhancement also improves the skeleton. Its weakness is that it hallucinates ridges on latents, and the teacher then rectifies a pattern that is not there. We run enhancement followed by the teacher as an additional baseline with matched latency accounting. If it equals the student on SD27 and HLD, the proposal reduces to a speed and simplicity argument, and we would say so.

Why not just use synthetic elastic warps or minutiae thin-plate splines instead of learned fields? The file's evidence argues against both as replacements. Sparse minutiae splines leave larger registration errors than learned dense methods: 5.65 to 5.95 pixels for classical dense methods against 3.75 for the network in Dense Registration and Mosaicking of Fingerprints by Training an End-to-End Network. Simple synthetic distortion models also miss real edge forces and pose effects. Our method keeps real, teacher-estimated geometry and synthesizes only appearance, which is the easier thing to simulate.

Why not just prune or quantize the teacher for deployment? Compression keeps the skeletonization dependency, and that dependency causes the failure on degraded prints. We report latency for both, but robustness is the main target.

Proposed Experiments

We will train the teacher and student on TDF with traced fields, plus the unlabeled clean pool, and evaluate on four sets: clean TDF_V2 (mean reprojection error), TDF_V2 under held-out corruption families that are never seen in training (mean reprojection error and matching false non-match rate), HLD Rank-1, and NIST SD27 Rank-1. All methods share one minutiae matcher.

The first decisive check is whether the student beats the rerun teacher on HLD and SD27, which are real latents and not synthetic corruptions. A gain only on synthetic degradation would mean the corruption model is the real contribution. The second check is whether the student beats the ground-truth-only ablation when both see the same corruptions. If it does not, the distillation claim fails. The third check is whether the student's advantage over the teacher grows as clean-skeleton quality on the test print falls. That pattern is what the mechanism predicts, and a flat curve would suggest the gain comes from something else. We also report GPU latency per 512×512 image, including the teacher's preprocessing time, against the published 0.41 seconds for the teacher network alone.