Abstract
One-to-many facial verification remains a significant challenge for mobile and edge computing due to the high computational demands of state-of-the-art models. This paper introduces CSEye, a lightweight, three-stage neural network architecture designed for efficient one-to-many verification. By utilizing a truncated VGG19 for feature extraction and employing weight-sharing across candidates, CSEye minimizes parameter count while maintaining high accuracy. Our results on the Labeled Faces in the Wild (LFW) dataset demonstrate that using angular distance for feature comparison yields a true positive rate of 98.74% in a 10-to-1 verification task, all while remaining trainable on standard CPU hardware.
Introduction
While one-to-one facial verification has reached a level of maturity that often surpasses human performance, one-to-many verification—identifying a "suspect" from a pool of $N$ candidates—presents unique hurdles. High-performing models typically rely on deep feature fusion and high-resolution imagery, which are often unavailable in mobile applications or low-quality surveillance contexts. This "practicality gap" necessitates a model that balances accuracy with computational efficiency. CSEye is proposed as a solution that prioritizes low-resolution processing and reduced architectural complexity to enable deployment in resource-constrained environments.
Methodology
The CSEye architecture is structured into three distinct stages: Feature Extraction, Comparison, and Classification.
Feature Extraction and Weight-Sharing
The model utilizes a modified VGG19 architecture as its backbone. To accommodate low-resolution inputs common in practical scenarios, the input size is reduced to $100 \times 100 \times 3$. The network is truncated after the final convolutional layer, removing the computationally expensive fully connected layers. This results in a feature map output of $512 \times 6 \times 6$. A weight-sharing paradigm is implemented to ensure that the suspect and all $N$ candidates are processed through identical extraction logic, significantly reducing the total parameter count.
Comparison Stage
The core innovation of CSEye lies in the explicit comparison of extracted features. Rather than concatenating features, the model calculates a relationship metric between the suspect's feature vector and each candidate's vector. Three primary metrics were evaluated:
- Angle: The cosine angle between vectorized feature maps.
- Dot Product: The scalar product of the vectors.
- Distance: The sum of element-wise differences.
Classification
The resulting comparison values are passed through a classification head consisting of three dense layers and a final softmax layer. This stage identifies the most likely match among the $N$ candidates by processing the relational data generated in the comparison stage.
Experimental Setup
Evaluation was conducted using the Labeled Faces in the Wild (LFW) dataset. To simulate one-to-many scenarios, custom samples were generated consisting of one "suspect" image, one "correct candidate" (a different image of the same individual), and nine "incorrect candidates" (randomly selected individuals). Testing was performed on an initial set of 5,000 samples and subsequently scaled to 10,000 samples to evaluate the model's generalization capabilities.
Results and Discussion
The choice of comparison metric proved critical to performance. In the initial 5,000-sample test, both the Angle and Distance metrics achieved 84% accuracy, outperforming the Dot Product (78%). Upon scaling to 10,000 samples, the Angle-based model demonstrated superior robustness, reaching a true positive rate (TPR) of 98.74%, compared to 90.35% for the Distance-based model.
The superiority of the Angle metric suggests that the orientation of feature vectors—representing the relative presence of specific facial traits—is a more reliable indicator of identity than raw feature magnitude. Furthermore, the model's efficiency is highlighted by its training performance; CSEye can be trained in under two seconds per epoch on a standard CPU, making it highly accessible for researchers and developers without expensive hardware stacks.
Conclusion
CSEye demonstrates that high-accuracy one-to-many facial verification is achievable without massive computational overhead. By focusing on truncated feature extraction and effective comparison metrics like angular distance, the model provides a portable solution for mobile and security applications. Future research will focus on scaling the candidate pool size ($N$) and fine-tuning the extraction layers to better capture subtle facial nuances.
Summary updated