Abstract
The COVID-19 pandemic and the resulting necessity for social distancing have increased the demand for reliable facial recognition systems in non-ideal conditions. However, the impact of the "glass factor"—the distortion caused by transparent barriers—remains under-researched. This paper provides a comprehensive review of how glass, water, and mist affect facial recognition accuracy. Through a two-step process of data collection and accuracy testing, we evaluate a state-of-the-art system utilizing MTCNN and Inception V1. Our findings indicate that while the system is robust when comparing two similarly disturbed images, it fails significantly when matching a real-time disturbed image against a clean database frontal image.
Introduction
The global pandemic normalized the use of glass partitions and face shields in high-traffic public areas such as airports and hospitals. While facial recognition technology has matured, most models are optimized for high-quality, unobstructed frontal images. This study identifies a critical gap: the refraction and distortion caused by glass, water droplets, and mist (the "glass factor") significantly impact the accuracy of these systems, particularly in "database-based" applications where a distorted real-time frame must be matched against a clear mugshot or ID photo.
Methodology
The study employs a standard two-stage facial recognition pipeline to evaluate the glass factor:
- Detection: The MTCNN (Multi-task Cascaded Convolutional Networks) framework locates faces. It operates through three stages—P-Net, R-Net, and O-Net—to propose candidate windows, refine them, and identify five facial landmarks.
- Recognition/Embedding: An Inception V1 deep network generates facial embeddings. The system calculates the Euclidean distance between these embeddings to determine identity.
Experimental Setup and Datasets
A custom dataset was curated to simulate real-world "disturbed" conditions. The dataset included 10 individuals (ages 4 to 84) across four distinct conditions: Normal (control), Clear Glass, Water Drop (simulating rain), and Mist (obscured by a nebulizer). A total of 120 images were captured, featuring three expressions per person per condition.
Key Results and Analysis
Detection Robustness
MTCNN proved remarkably resilient to environmental noise. It achieved a 95% success rate in extracting faces even through water and mist, failing in only 3 out of 60 disturbed cases. This suggests that the initial detection phase is not the primary bottleneck for glass-factor interference.
Recognition Thresholds and the "Database Gap"
The core challenge lies in the embedding distance. The "glass factor" significantly increases the Euclidean distance between embeddings of the same person (self-distance).
- Real-time Comparison: The system is relatively effective at comparing two "disturbed" images (e.g., Glass to Glass), where a threshold of 0.4 successfully distinguished identities.
- Database Comparison: The system failed the "Normal to Water" and "Normal to Mist" tests. In these scenarios, the refraction acts as non-linear noise, making the subject appear more different from their own clear database photo than from a different person. For example, in one test, the self-distance rose to 1.05, while the distance to a different person was only 0.99, leading to a false rejection.
Conclusion
While modern facial recognition architectures like MTCNN and Inception V1 are robust in detection, they struggle with the refractive noise introduced by water and mist during the recognition phase. Future work should focus on fine-tuning models specifically on glass-factor datasets. By exposing neural networks to these specific refractive patterns during training, the system may learn to ignore these artifacts, closing the performance gap between real-time disturbed images and clean database records.
Summary updated