PaperlayerView exact revision on arXiv

Source-linked AI summary

Balancing Foundation Models and Trait-Specific Learning for Forehead-Crease Recognition

Geetanjali Sharma, Aryan Jain, Dev Nagaich, Aditya Nigam, Raghavendra Ramachandra

arXiv:2610.29495v1cs.CV

TL;DR

Partial facial occlusion weakens holistic face recognition, while forehead-crease recognition has lacked unified comparisons across model paradigms. The paper evaluates explainable zero-shot reasoning, domain-adaptive learning, and trait-specific modeling; trait-specific and domain-adaptive systems outperform MLLMs, though MLLMs provide interpretable zero-shot verification.

  • Problem

    Partial facial occlusion limits holistic face representations, and forehead-crease research has limited pretrained-model use, inconsistent protocols, and insufficient study of explainability and domain shift.

    Source: §1 · p. 1, §2.4 · p. 5

  • Method

    The paper evaluates explainable multimodal zero-shot reasoning, margin-based adaptation of pretrained face-recognition models, and a trait-specific crease model under unified open-set protocols.

    Source: §1 · p. 2, §3 · p. 5

  • Results

    Trait-specific and domain-adaptive systems outperform MLLMs; Edge-Forehead achieves the lowest EER (8.30%).

    Source: Quantitative Results · p. 13, §6 · p. 14

  • Takeaways & Limitations

    Trait-specific learning remains more effective for robust forehead-crease recognition, while MLLMs offer interpretable zero-shot verification.

    Source: §6 · p. 14

    Generic foundation models are not designed for subtle crease cues, and deployment raises computational-cost, latency, and scalability challenges.

    Source: §2.3 · p. 5

Abstract

from arXiv · show

Biometric recognition often degrades under partial facial vis- ibility, where conventional facial cues are occluded. The forehead region remains consistently visible and contains distinctive crease patterns that can serve as a complementary biometric trait. We present a unified eval- uation of forehead-crease recognition across explainable zero-shot rea- soning, domain-adaptive learning, and trait-specific modeling. FIRE en- ables interpretable zero-shot verification using multimodal large language models, while pretrained face-recognition models are adapted to the fore- head domain through margin-based learning. Under open-set evaluation, the best zero-shot model GPT-4o+FIRE achieves 18.22% EER, while Edge-Forehead achieves 8.30% EER and 54.01% TMR@0.1% FMR. The results show that although MLLMs offer interpretable zero-shot verifi- cation, trait-specific learning remains more effective for robust forehead- crease biometric recognition.

1 Introduction

Forehead creases offer identity cues when faces are partially occluded, but their recognition remains underexplored across transferable models, fair protocols, and explainability. The paper unifies zero-shot reasoning, domain adaptation, and trait-specific modeling for open-set evaluation.

1 Introduction

  • Forehead creases retain identity-discriminative information when global facial cues are unavailable, and the forehead is consistently visible.Prior studies demonstrated verification across images from two sessions, but temporal stability remains insufficiently quantified.Source: §1 · pp. 1–2
  • Existing forehead-crease methods often lack transferable face representations and fair cross-paradigm protocols, while explainability remains underexplored.The paper asks whether pretrained face-recognition models can be adapted to the spatially constrained forehead domain.Source: §1 · p. 2
  • The proposed framework combines explainable multimodal zero-shot reasoning, margin-based domain adaptation, and explicit modeling of fine-grained crease structures.It evaluates these paradigms under open-set protocols on real-world and synthetic forehead datasets.Source: §1 · p. 2
  • Forehead-crease biometrics can complement recognition in partial-face scenarios without additional sensors or intrusive acquisition.The paper also contributes FIRE for interpretable multimodal verification and an open-set comparison of the three paradigms.Source: §1 · p. 3

2 Related work

The review frames forehead-crease recognition around limited transfer from face-recognition models, inconsistent evaluation, and underexplored explainability and domain shift. It surveys forehead-specific methods, domain adaptation, and multimodal verification before motivating a unified comparison.

2 Related work

  • The review covers forehead-based biometrics, face-recognition and domain adaptation, and explainable verification with multimodal models.It positions the paper against limitations across these research directions.Source: §2 · p. 3

2.1 Forehead-Based Biometric Recognition

  • Forehead methods progressed from handcrafted descriptors to crease-specific networks and synthetic data, but synthetic-trained models can degrade on real images.Earlier work modeled wrinkle patterns and skin structure; newer approaches learn crease arrangement, curvature, and local texture.Source: §2.1 · pp. 3–4

2.2 Face Recognition and Domain Adaptation

  • Pretrained face models and margin-based losses may transfer discriminative representations to the forehead, but this trait-specific transfer remains underexplored.AdaFace and CosFace encourage intra-class compactness and inter-class separability.Source: §2.2 · p. 4

2.3 Multimodal Large Language Models

  • Multimodal models offer zero-shot, human-interpretable verification, but general-purpose models are not designed to resolve subtle crease patterns.Their predictions can also vary with illumination, resolution, and expression, and may include hallucinated reasoning when samples are similar.Source: §2.3 · pp. 4–5

2.4 Limitations of Existing Work

  • The paper addresses limited pretrained-model use, inconsistent protocols, and insufficient study of explainability and domain shift through a unified evaluation.The comparison considers interpretability, transferability, and verification performance.Source: §2.4 · p. 5

3 Proposed Method

The method integrates explainable zero-shot verification, adaptation of pretrained face-recognition models, and a trait-specific crease model. It formulates forehead-image comparison as binary identity verification with a similarity score and, for explainable models, a textual rationale.

3 Proposed Method

  • The framework combines multimodal zero-shot reasoning, domain adaptation of pretrained face-recognition networks, and trait-specific modeling of crease structures.The pipeline is designed to operate solely on forehead regions when full-face information is unavailable.Source: §3 · p. 5

3.1 Problem Formulation

  • A pair of forehead images is treated as a binary task: determine whether both images belong to the same identity.The target is 1 for images of the same subject and 0 otherwise.Source: §3.1 · p. 5
  • The system returns a similarity score, with explainable models additionally producing a structured textual explanation of their decision.The score is defined on the interval [0, 1].Source: §3.1 · p. 6

3.2 Forehead Region Extraction and Preprocessing

The preprocessing pipeline standardizes forehead crops for model input while applying training-time perturbations intended to improve generalization. The overview figure depicts evaluation of real and synthetic image pairs across model families.

3.2 Forehead Region Extraction and Preprocessing

  • Images are cropped to the forehead with a fixed strategy, converted to grayscale, replicated across three channels, and resized to 224 × 224.This provides spatial consistency and compatibility with standard CNN architectures.Source: §3.2 · p. 6
  • Training augmentation uses horizontal flips, small rotations, and brightness perturbations while preserving discriminative crease patterns.These transformations are applied to improve generalization.Source: §3.2 · p. 6
  • The pipeline figure depicts real and synthetic image pairs evaluated with multimodal, vision foundation, and trait-specific models.It summarizes the evaluated model families rather than reporting a performance comparison.Source: Figure 1 · p. 6

3.3 Explainable Zero-Shot Verification via FIRE

FIRE structures multimodal models’ forehead-image comparisons into interpretable verification decisions. Vision foundation models provide zero-shot baselines, but their performance is sensitive to image and domain variation.

3.3 Explainable Zero-Shot Verification via FIRE

  • FIRE guides multimodal models through image-condition assessment, crease-feature extraction, feature comparison, and a confidence-based match decision.It returns a classification and human-readable explanation; the paper says structured reasoning reduces hallucinations and improves consistency versus generic prompts.Source: §3.3 · p. 7

3.4 Zero-Shot Verification using Vision Foundation Models

  • DINOv2 and CLIP are evaluated without task-specific fine-tuning using embeddings and cosine similarity for verification.Source: §3.4 · p. 7
  • Zero-shot foundation models are sensitive to illumination, expression, and domain shift because they are not trained specifically for forehead biometrics.DINOv2 captures fine-grained structural features, while CLIP benefits from image-text semantic alignment.Source: §3.4 · p. 7

3.5 Domain-Adaptive Learning using Pretrained Models

The approach transfers pretrained face-recognition representations to the forehead domain and refines them through fine-tuning for forehead-specific features.

3.5 Domain-Adaptive Learning using Pretrained Models

  • The domain-adaptive strategy transfers knowledge from pretrained face-recognition models to the forehead domain.Source: §3.5 · p. 7

Feature Representation

Margin-Based Metric Learning

Margin-based objectives are used to make forehead-crease representations more discriminative, while FIRE specifies a structured process for comparing forehead images.

Margin-Based Metric Learning

  • Margin-based losses enforce compact intra-class distributions and well-separated inter-class boundaries.Source: Margin-Based Metric Learning · p. 8
  • The evaluated objectives are ElasticFace, AdaFace, MagFace, and CosFace, with different margin strategies for discriminative forehead-crease modeling.The passage characterizes them respectively as stochastic margin regularization, quality-adaptive margin, magnitude-aware representation learning, and additive cosine margin.Source: Margin-Based Metric Learning · p. 8

Forehead Interpretation and Reasoning Engine (FIRE)

  • FIRE assesses image quality, compares crease characteristics, accounts for variations, and produces a match decision with confidence, dissimilarity, and reasoning.It constrains decisions to observable forehead features and requires uncertainty to be reported when evidence is insufficient.Source: Forehead Interpretation and Reasoning… · p. 8

3.6 Trait-Specific Model: Edge-Forehead

Edge-Forehead is a trait-specific model that fine-tunes EdgeFace for forehead-crease recognition, combining gradual unfreezing with a dual-loss objective.

3.6 Trait-Specific Model: Edge-Forehead

  • Edge-Forehead fine-tunes EdgeFace on forehead images with gradual unfreezing to retain general facial features while specializing in crease patterns.Source: §3.6 · p. 8
  • The dual-loss objective combines Lce for inter-class separability and Lcenter for reducing intra-class variance.Parameters α and β balance classification accuracy and feature compactness.Source: §3.6 · p. 8
  • Forehead-only modeling targets crease continuity, spacing, curvature, and texture patterns, and the models are compared under a unified verification protocol.The passage reports improved robustness under illumination, pose, and expression variations.Source: §3.6 · p. 8

4 Dataset & Evaluation Protocol

The evaluation uses real and synthetic datasets for zero-shot verification and a subject-disjoint FH-V1 training set with open-set testing on unseen FH-V2 identities. A unified protocol compares zero-shot, domain-adaptive, and trait-specific systems.

4 Dataset & Evaluation Protocol

  • The protocol evaluates forehead creases as the primary biometric trait across real and synthetic data, with a unified comparison spanning MLLMs, VFMs, and supervised systems.The section frames this setup as enabling fair comparison across model types.Source: §4 · pp. 8–9

4.1 Datasets for Zero-Shot Evaluation (LLMs & VFMs)

  • Zero-shot evaluation samples 50 genuine pairs from 50 subjects and 450 impostor pairs from 250 subjects in FH-V2 and the synthetic dataset.The pairs capture variation in illumination, expression, and image quality.Source: §4.1 · p. 9

4.2 Datasets for Domain-Adaptive & Trait-Specific Models

  • Supervised models train on FH-V1's 4,964 images from 247 subjects and are tested open-set on FH-V2's 295 unseen subjects.FH-V2 contains 29,500 genuine and approximately 8.67 million impostor pairs; subject-disjoint training and validation prevent identity leakage.Source: §4.2 · p. 9

4.3 Training Protocols

  • Domain-adaptive backbones use margin-based objectives, while Edge-Forehead follows a three-stage schedule combining classification, center loss, and semi-hard triplet mining.Training details include progressive unfreezing and augmentation; Edge-Forehead uses 512-dimensional embeddings for verification.Source: §4.3 · pp. 9–10

4.4 Evaluation Protocol

  • Models are compared under open-set verification using EER and TMR at 0.1% and 0.01% FMR operating points.Zero-shot decisions use majority voting across three stateless runs; reproducibility settings include a subject-disjoint 90%/10% split with seed 42.Source: §4.4 · p. 10

5 Results & Discussion

The results compare zero-shot, domain-adaptive, and trait-specific forehead-crease verification under a unified evaluation. Trait-specific learning achieves the strongest full open-set FH-V2 result, while zero-shot models provide interpretable verification and show response variability.

5.1 FIRE-based LLM and VFM Results

  • Among zero-shot methods, GPT-4o achieves the strongest reported performance (18.22% EER, 52.00% TMR); on synthetic data, it remains best among zero-shot models (25.44% EER) but struggles with synthetic variations.Source: Quantitative Analysis · p. 11

Qualitative Analysis

  • Across three inference rounds, FIRE responses show refinement, stable high-confidence decisions, recovery from refusals, and variability on challenging cases.The authors attribute conflicting predictions in difficult cases to model sensitivity and stochasticity.Source: Qualitative Analysis · p. 12, Table 2 · p. 12

Explainability Analysis

  • FIRE pairs verification decisions with human-interpretable explanations, including crease-alignment and spacing differences for impostor pairs.Examples report up to 85% confidence for impostor pairs and 95% for genuine matches, while noting occasional errors under challenging visual conditions.Source: Explainability Analysis · p. 12

Quantitative Results

  • MobileFaceNet-V2 with CosFace achieves the lowest domain-adaptive EER at 9.31%, while FaceNet with ElasticFace reaches 11.17% and SwinFace with AdaFace reaches 13.44%.EdgeFace-Xs with MagFace achieves 11.01% EER; the findings indicate that backbone architecture and margin-based supervision both matter.Source: Quantitative Results · pp. 12–13, Table 3 · p. 13
  • 8.30% EER and 70.00% TMR at both operating points: Edge-Forehead leads the full open-set FH-V2 comparison, ahead of GPT-4o's 18.22% EER.Edge-Forehead also surpasses FHCVS (21.56% EER) and FH-SSTNet (14.22% EER).Source: Quantitative Analysis · p. 11, Table 3 · p. 13, Quantitative Results · p. 13
  • Edge-Forehead achieves the lowest EER in the trait-specific comparison, while MFN-V2+CosFace provides the strongest low-FMR performance.The DET curves show the open-set FH-V2 verification comparison; SwinFace variants have higher FNMRs.Source: Quantitative Results · pp. 13–14, Figure 2 · p. 14

6 Conclusion

The paper finds that domain-adaptive and trait-specific models outperform MLLMs in forehead-crease verification, while MLLMs offer interpretable zero-shot verification.

6 Conclusion

  • Domain-adaptive and trait-specific models consistently achieve lower EER and higher TMR than MLLMs, supporting trait-specific learning for reliable security-critical biometric deployment.The paper's comparison covers GPT-4o, Claude, Gemini, domain-adaptive learning, and trait-specific models.Source: §6 · p. 14
Loading 2610.29495v1…