anbinh39/LFR-GAN

G_render — identity-preserving latent-fingerprint enhancement

g_render2 renders a clean, examiner-friendly fingerprint image from a degraded latent while preserving identity (ridge angles + minutiae positions). It is a conditional inpainter with a window-attention U-Net backbone and a hard coverage-gate that forbids inventing ridges where the input has no evidence — the mechanism that keeps enhancement from destroying identity.

• Backbone: RenderGenerator = UformerCondGenerator (window spatial self-attention U-Net), 2.94M params
• Input: 6-channel conditioning at 256×256 → Output: 1-channel enhanced image
• Trained: supervised degrade(clean) → clean on NIST SD302 (1999 rolled + 2000 plain prints)
• Benchmark: NIST SD302, 487 latent probes / 1999 rolled gallery

Results (ALL 487 probes)

Image quality — best of all methods tested:

Method	RQI ↑	KID ↓	NFIQ2 ↑
g_render2 (this)	0.811	0.165	17.1
clan_efsroi	0.672	0.197	11.1
FLARE	0.624	0.165	14.2
FingerGAN	0.393	0.264	36.8
raw latent	0.290	0.263	8.3

Identity (FingerNet → DMD dense-descriptor matcher):

Method	Rank-1 ↑	AUC ↑	EER ↓
g_render2 + classic_lan post-proc	0.610	0.799	0.232
raw latent	0.608	0.796	0.244
g_render2	0.600	0.788	0.261
FLARE	0.559	0.749	0.296
FingerGAN	0.353	0.722	0.335

g_render2 is the only deep enhancer that reaches best-in-class image quality while keeping matcher identity at or above the raw latent (all published deep enhancers we tested fall below the raw latent on the matcher). Adding a 0-param analytic classic_lan post-process nudges identity slightly above the raw latent (Rank-1 0.610 ≥ 0.608).

Usage

python inference.py input_latent.png output_enhanced.png \
    [--roi roi_mask.png] [--minu minutiae_xy.txt] [--device cuda]

import torch
from g_render.models.render_generator import RenderGenerator
from inference import enhance

model = RenderGenerator(in_ch=6).eval()
model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"), strict=False)
enhanced = enhance(model, "input_latent.png", roi_path=None, minu=None, device="cpu")  # PIL.Image
enhanced.save("output.png")

Input conditioning (6 channels, built by inference.build_cond)

ch	content
0	input image (classic_lan-normalized) — the only raw-pixel channel
1–2	robust orientation field [cos2θ, sin2θ] · coverage
3	coverage / ROI mask
4	ridge-frequency map · coverage
5	minutiae Gaussian heatmap

--roi (foreground mask) and --minu (minutiae) are optional: without them the ROI is derived from ridge coherence and the minutiae channel is zero. A curated ROI (e.g. a learned segmenter or an examiner-marked region) gives the cleanest result — the coverage-gate blanks everything outside it.

Files

pytorch_model.bin (weights) · config.json · inference.py (full pipeline) · g_render/ (self-contained model code, no external repo needed).

Limitations

• Does not add identity information (data-processing-inequality ceiling): it preserves the latent's identity, it does not exceed it. For the strongest automated matching, the analytic classic_lan baseline still leads (Rank-1 0.706); use g_render2 when a clean, realistic image that keeps identity is the goal (human examiners, visualization).
• Trained/evaluated on 500-ppi NIST SD302 latents; other sensors/resolutions may need re-tuning.
• Minutiae positions are jittered slightly by the re-render (weaker on the position-only Bozorth3 matcher than on the dense DMD matcher).

Citation / provenance

Derived from the FDC-LFE project. Bench protocol: FingerNet→DMD / FingerNet→Bozorth3 on NIST SD302 (487 probes / 1999 gallery). Post-processing baseline classic_lan = local adaptive contrast norm (target_std 0.18, gain∈[0.5,4], k=13). License follows the NIST SD302 data terms — verify before use.