Sparse Linear Polarization Pixels for Face Anti-Spoofing

Abstract

Overview: Hybrid sensor with sparse polarization pixels for face anti-spoofing

Robust face anti-spoofing (FAS) is essential for secure facial recognition systems. This work presents a novel hybrid sensor approach using sparsely integrated linear polarization pixels into an RGB pixel matrix to leverage both angle of linear polarization (AoLP) and degree of linear polarization (DoLP). By integrating polarization features into a lightweight CNN, our solution offers a cost-effective and reliable FAS under all light conditions.

                    Key Achievement: We achieve an average classification error rate (ACER) as low as 0.4%, even with an extremely sparse deployment of polarization pixels (1 per 400 RGB pixels), representing a 10-fold reduction in errors compared to RGB-based methods.
                

Key Results

0.4%

ACER (Best Performance)

0.25%

Polarization Pixel Density

10×

Error Reduction vs RGB

99.97%

AUC ROC

Method Overview

Figure 3: End-to-End Inference Pipeline - from polarization image capture to FAS classification

Hybrid Sensor Design

Our approach uses a hybrid sensor that sparsely integrates polarization pixels alongside standard RGB pixels. This design:

Minimizes RGB image degradation (only 0.25% of pixels used for polarization)
Reduces computational complexity
Maintains high-quality RGB imaging for face recognition
Enables robust FAS even under low-light conditions (100 lux)

Linear Polarization Colormap

We introduce a novel three-channel representation combining:

AoLP (Angle of Linear Polarization): Captures 3D facial geometry through Cartesian coordinates
DoLP (Degree of Linear Polarization): Quantifies polarization intensity

                    Our ablation study reveals that AoLP is more informative than DoLP for FAS, contrary to previous assumptions. However, combining both provides the best performance.
                

Performance Comparison

Method	Modality	ACER (%)	AUC ROC (%)
Sparse Polarization (Ours)	Polarization (0.25% pixels)	0.39	99.97
Dense Polarization (Ours)	Polarization (100%)	0.32	99.97
RGB Baseline	RGB	3.77	98.79
RGB + Depth (ViT+AMA)	RGB + Depth	1.65	N/A
Multi-modal (MCCNN)	RGB + Depth + NIR + Thermal	0.2	N/A

Ablation Study: Impact of Input Representation

Figure 6: Equal Error Rate (EER) as a function of training set size for different input representations. Our Linear Polarization Colormap (combining AoLP and DoLP) consistently outperforms single-modality approaches. AoLP alone is more informative than DoLP alone.

Key Finding from Ablation Study

For any required accuracy level, methods using only DoLP or AoLP Grayscale require exponentially more training data compared to our Linear Polarization Colormap representation. AoLP contributes more discriminative information than DoLP, but combining both yields the best results.

Key Contributions

✓ First demonstration of effective FAS using extremely sparse polarization pixels (1 per 400 RGB pixels)
✓ Novel linear polarization colormap representation combining AoLP and DoLP
✓ Comprehensive ablation study showing AoLP is more informative than DoLP for FAS
✓ Robust performance under low-light conditions (100 lux)
✓ Lightweight CNN architecture (MobileNetV3-Small) suitable for mobile devices
✓ New dataset: Polarized FAS Dataset (PFASD) with 64 subjects under controlled and realistic lighting

Visual Results

Polarization Images: Genuine vs Spoofing Attacks

(a) Print attack with minimal variation

(b) Genuine face with polarization-dependent highlights

(c) Replay attack with pronounced contrast variations

Figure 2: Polarization images at different angles (0°, 45°, 90°, 135°) showing distinct characteristics for each presentation type

Linear Polarization Colormap Representation

Linear Polarization Colormap visualization

Figure 4: Our three-channel Linear Polarization Colormap at different sparsity levels. Red/green channels encode AoLP, blue channel represents DoLP. Shows genuine face, print spoof, and replay spoof at dense, sparsity 10, and sparsity 20 levels.

Low-Light Performance

(a) RGB image appears very dark

(b) Polarization colormap clearly reveals 3D facial structure

Figure 5: Visual comparison under 100 lux low-light conditions showing the superior performance of polarization imaging

Impact & Applications

Why This Matters

This work enables secure, affordable face authentication for mobile devices and embedded systems by:

Using a single hybrid sensor instead of multiple expensive sensors
Maintaining RGB image quality for face recognition
Requiring minimal computational resources (lightweight CNN)
Working reliably even in challenging lighting conditions

Citation

@article{kim2025sparse,
  title={On the Effectiveness of Sparse Linear Polarization Pixels for Face Anti-Spoofing},
  author={Kim, JaeSeong and Pelz, Abraham and Scherer, Michael and Mendlovic, David},
  journal={IEEE Sensors Journal},
  volume={25},
  number={18},
  pages={35178--35190},
  year={2025},
  publisher={IEEE},
  doi={10.1109/JSEN.2025.3597155}
}