TL;DR
This paper introduces forehead creases as a novel biometric modality for mobile-based human identification, particularly relevant in COVID-19 scenarios where face masks occlude traditional facial recognition. They propose a deep learning framework with a dual attention mechanism and a new dataset of masked forehead images, achieving high accuracy in both closed-set and open-set recognition tasks.
What Problem Does This Paper Address?
The COVID-19 pandemic has significantly challenged existing biometric authentication systems. Face masks obscure a large portion of the face, rendering face recognition unreliable. Furthermore, the need to avoid physical contact has made fingerprint scanners less desirable due to hygiene concerns. While iris scanners are an option, they often require high-quality images and can be incompatible with spectacles. This paper addresses the urgent need for a contactless, reliable, and mask-compatible biometric modality for mobile devices. They propose utilizing forehead creases, which are formed by voluntary facial expressions, as a unique and persistent biometric trait.
Methods
The proposed framework is an end-to-end deep learning system for forehead recognition. Key components include:
- Forehead ROI Segmentation: A YOLOv3 model, pre-trained on ImageNet and fine-tuned on forehead images, is used to extract a Region of Interest (ROI) from the forehead area. This step is crucial for handling variations in pose, scale, and illumination.
- Backbone Network: A ResNet-18 architecture is employed as the backbone to extract initial feature maps. ResNet-18 was chosen for its efficiency and proven performance on large-scale datasets, aiming to prevent overfitting on their relatively smaller dataset.
- Dual Attention Module (DAM): This novel component enhances feature discriminability. It comprises:
- Spatial Attention Module (SAM): Aggregates semantically similar pixels to capture spatial variations in lines and wrinkles.
-
Channel Attention Module (CAM): Focuses on essential feature maps and reduces redundancy by modeling channel-wise relationships. The outputs from SAM and CAM are fused using either concatenation or element-wise addition to create a richer feature representation.
-
Loss Function: Large Margin Cosine Loss (LMCL) is used to learn discriminative features in a cosine space, promoting intra-class similarity and inter-class separation. This is further optimized with Focal Loss for better weight updates.
- Data Acquisition Strategy: A custom Android application facilitates remote, non-contact data collection, allowing users to capture forehead images with specific poses and expressions designed to reveal creases.
Datasets
The paper introduces a new dataset: BITS-IITMandi-ForeheadCreases Images Database.
- Size: Contains 4,964 selfie-face mask images.
- Subjects: Collected from 247 subjects.
- Collection: Recorded in two sessions, with subjects providing images from two poses in each session. This temporal separation helps assess the stability of forehead patterns.
- Augmentation: The dataset was augmented by 20 times for training Model V2, which involved sampling points around the original bounding box to create new ROIs.
Key Insights and Ideas
- Forehead Creases as a Biometric: The core idea is that forehead creases, similar to finger knuckle prints or palm lines, possess unique and persistent patterns that can be leveraged for identification.
- Dual Attention for Feature Enhancement: The combination of spatial and channel attention is shown to be more effective than individual attention mechanisms or no attention at all, leading to more robust feature learning.
- Importance of Data Augmentation: Training with augmented data (Model V2) significantly improves performance, especially in open-set scenarios, highlighting the network's ability to generalize to unseen variations.
- Efficiency for Mobile Deployment: The framework is designed to be efficient, requiring minimal hardware and memory, making it suitable for mobile applications.
Key Results
The proposed system achieved impressive results:
- Closed-Set Matching:
- Model V2 (trained on augmented data) achieved a Correct Recognition Rate (CRR) of 99.08% and an Equal Error Rate (EER) of 0.44%.
- Open-Set Matching:
- Model V2 achieved a CRR of 97.84% and an EER of 12.40%.
These results are comparable to established biometrics like iris and finger knuckle recognition. The ablation study also indicated that the dual attention mechanism, particularly with element-wise sum fusion, provided superior performance.
Limitations and Future Work
- Limitations:
- The dataset, while new, is still relatively small for deep learning standards, especially for training from scratch.
- The performance in open-set matching, while good, is naturally lower than closed-set matching, indicating challenges with truly unseen subjects.
- The paper acknowledges that the "full potential from the forehead biometrics is yet to be realized" due to challenges like pose, scale, and illumination variations in unconstrained environments.
- Future Work:
- The authors suggest further research into improving the robustness of the system to more challenging unconstrained environments.
- Exploring fusion with other biometrics could enhance overall system accuracy and security.
- Investigating the temporal stability of forehead creases over longer periods could further validate their long-term reliability.
Summary updated