August 2026 Face Recognition Papers: Low-Resolution Deployment, Sub-1 kB Compression, and Brazilian Fairness

August's face recognition papers focus on constraints that become visible only after a model leaves a clean benchmark. The selected studies test native low-resolution probes rather than trusting synthetic degradation, quantify which codecs preserve identity at 1,024 and 512 bytes, and audit compressed public video with Brazilian demographic categories and explicit difficulty strata.

What This Month Signals

The low-resolution study warns that the synthetic degradation that wins on LFW-style tests can be the wrong choice for native TinyFace probes. The compression benchmark shows a similarly sharp regime change: several familiar codecs work near 1 KB but collapse at 512 B, where a byte-budgeted learned codec becomes valuable. UFPR-PEs then demonstrates that demographic gaps grow dramatically in the hard visual subset and can change ordering between identification protocols. Across all three, production quality depends on testing the actual capture and storage regime rather than extrapolating from a convenient proxy.

Paper 012026-08-06cs.CV

Improving Low-Resolution Face Recognition under Limited Data: How Synthetic Data Generation Can Close the Domain Gap

arXiv PDF

Authors & Institutions

Luis S. Luevano

Idiap Research Institute, Switzerland

Ünsal Öztürk

Idiap Research Institute, Switzerland

Hatef Otroshi Shahreza

Idiap Research Institute, Switzerland

Anjith George

Idiap Research Institute, Switzerland

Sébastien Marcel

Idiap Research Institute, Switzerland

What Problem It Solves

The study asks which synthetic low-resolution strategy actually helps a compact face recognizer when labeled native data is limited, and whether conclusions drawn from artificially downsampled LFW-style tests survive on real TinyFace probes. It also establishes a direct-input baseline so that super-resolution and translation modules must prove they preserve identity rather than merely improve visual appearance.

Key Result

Synthetic tests favored a 28-pixel degradation, but that same setting fell below the high-resolution-trained baseline on real TinyFace. Milder 56-pixel interpolation augmentation was the best compact configuration, raising TinyFace mAP from 52.55 to 54.68 and rank-1 identification from 59.66% to 61.40%. The learned RRDB super-resolution plus domain-translation pipeline reached only 57.53 mAP with EdgeFace-base, versus 65.12 for direct input. Knowledge distillation delivered the largest synthetic gains but did not transfer to native low resolution, showing that synthesis fidelity, not synthesis complexity, is the limiting factor.

Abstract

Face Recognition (FR) systems in surveillance settings often encounter Low Resolution (LR) faces, those whose face region falls below the standard 112 $\times$ 112 input size. While labelled High Resolution (HR) training data is abundant, labelled native-LR data, and above all paired native LR/HR data, is scarce. One workaround is to synthesize LR data from the available HR faces, but how much synthesis effort is repaid in recognition accuracy remains unclear. We present a study of simple synthetic generation strategies for a compact, edge device-oriented face recognition system, spanning interpolation-based degradation, knowledge distillation, a Prepended Domain Transformer (PDT), Real ESRGAN-style degradation, and a learned Super Resolution (SR) front-end with an identity-aware loss. We evaluate these strategies on synthetic cross-resolution face benchmarks (LFW, CFP-FP, AgeDB-30) and on TinyFace, a real-world native LR dataset, and expose a synthetic-real gap: the degradation setting that is optimal on synthetic benchmarks is not the one that is optimal on real LR. We find that more synthesis effort does not help monotonically: the learned SR front-end does not surpass a direct feed of the aligned LR image into a strong backbone, while simple interpolation augmentation of a compact backbone is the only synthesis that improves over its own baseline. We conclude that generative methods for LR face recognition must be validated on real LR and against a direct-feed baseline, and release our pipeline at https://idiap.ch/paper/synth-lrfr

Research Starting Point

Surveillance and watchlist systems routinely compare a high-quality enrollment image with a probe whose aligned face is far smaller than the standard 112 by 112 input. Native low-resolution identities, especially paired low- and high-resolution captures of the same person, are scarce, so teams commonly manufacture low-resolution training data from abundant high-resolution faces. The practical risk is that a degradation recipe can score well on synthetic benchmarks while teaching the recognizer cues that do not resemble blur, noise, compression, and alignment errors from real cameras.

Method

The authors use the 3.65-million-parameter EdgeFace-S backbone and compare five levels of adaptation effort: interpolation-based downsample-and-upsample augmentation, high-resolution teacher knowledge distillation, a prepended domain transformer, a native 32-pixel stem trained with Real-ESRGAN-style degradations, and an identity-aware learned super-resolution front end. They evaluate high-to-low and low-to-low verification on LFW, CFP-FP, and AgeDB-30, then repeat the comparison on native low-resolution TinyFace under two alignment pipelines. A larger frozen EdgeFace-base model provides the reference for testing whether a generative front end beats feeding the aligned probe directly to a strong recognizer.

Paper Summary

For an edge or surveillance deployment, validate every low-resolution recipe on genuinely captured low-resolution faces and keep a strong direct-feed baseline. Simple, mild interpolation augmentation can be a better investment than a super-resolution stack; synthetic-only gains at aggressive resolutions are not reliable evidence of field performance.

Paper 022026-08-24cs.CV

Toward Sub-1 kB Identity-Preserving Face Compression: A Benchmark of Codecs, a Custom Learned Codec, and Studies of Resolution, Demographic Fairness, Recompression, and Adversarial Robustness

arXiv PDF

Authors & Institutions

Petr Hurtik

Innovatrics, Slovakia

Jakub Sochor

Innovatrics, Slovakia

What Problem It Solves

This work turns sub-kilobyte face storage into a controlled deployment benchmark instead of a single codec comparison. It identifies budget-specific operating points, tests whether image-quality metrics predict biometric utility, and builds a learned codec whose output is guaranteed to stay within the requested byte limit.

Key Result

At 1,024 bytes and 112 pixels, modern codecs are close to operationally solved on Color FERET: WebP and AVIF reach 0.09% EER with ArcFace. At 512 bytes, however, AVIF, HEIF, JPEG XL, and legacy JPEG rise to 28-98% FNMR at FMR 1e-4, while the accurate learned codec reaches 1.83% on Color FERET and 6.9% on AI-Solutions-KK, compared with 2.86% for JPEG-AI and 24.3% for WebP respectively. Codec order is largely backbone-invariant (Kendall's W=0.85). Most codecs add no more than 1.7 percentage points of demographic EER gap, but JPEG 2000 adds 3.4-5.0 points and also preserves identity poorly.

Abstract

Storing face images under a hard sub-kilobyte budget, as required for identity documents, smart-card biometrics and bandwidth-constrained verification, forces a codec to discard most of the signal while keeping what a face matcher actually reads: identity. Generic codecs optimize pixel fidelity, not the embedding distances that drive verification, so which codec, resolution and setting best preserve identity at 1024 bytes or less, and how that degrades at 512, is unclear. We benchmark ten general and face-specific codecs across resolutions, byte budgets, two datasets (controlled Color FERET, in-the-wild AI-Solutions-KK) and four anchor face matchers, with a fourteen-model ViT and CNN roster confirming the ranking is backbone-invariant. We then train a custom identity-preserving codec that hits the byte budget exactly via binary search over a frozen gain table, and run four studies: resolution, demographic fairness, recompression, and no-box adversarial robustness. Sub-kilobyte identity preservation is feasible, but which codec to deploy depends entirely on the budget. At 1024 bytes and the 112 px working resolution the problem is close to solved: modern codecs hold Color FERET equal-error rate under 0.35 percent on the ArcFace anchor. At 512 bytes the field re-sorts: AVIF, HEIF, JPEG XL and legacy JPEG collapse to 28 to 98 percent false-non-match rate at FMR 1e-4, while WebP, JPEG-AI and our byte-budgeted learned codecs stay out of that band, with 24.3 percent for WebP against 6.9 percent for our accurate variant in the wild. That re-sort, not the 1024-byte ranking, is the operational result: a codec chosen at 1 kB is not the codec to deploy at half that.

Research Starting Point

Identity documents, smart cards, and bandwidth-constrained verification services may have only 1,024 or even 512 bytes for an aligned face crop. Generic codecs optimize pixel distortion rather than the embedding distances used by a matcher, and a codec that looks best at one budget may fail abruptly at another. Buyers also need to know whether the choice transfers across recognizers, affects demographic groups, survives recompression, and interacts with adversarial perturbations.

Method

Ten general-purpose and face-oriented codecs are tested at five resolutions and two hard budgets on controlled Color FERET and in-the-wild AI-Solutions-KK. Four anchor matchers carry the main verification metrics, while a fourteen-model ViT/CNN roster checks whether codec rankings depend on the backbone. The authors also train 18.7-million-parameter accurate and fast identity-preserving codecs with a frozen 64-entry gain table and binary search for exact budget compliance. Separate studies cover skin-tone and ethnicity disparity, same- and cross-codec recompression, no-box adversarial perturbations, latency, and statistical significance.

Paper Summary

Treat 1 KB and 512 B as different product tiers, not neighboring quality settings. WebP or AVIF are sensible broadly deployable choices around 1 KB, while a byte-budgeted learned codec becomes materially safer at 512 B; biometric tests, not PSNR or visual inspection alone, should drive the decision.

Paper 032026-08-31cs.CV