The emergence of face masks, first due to the pandemic and later as a requirement in high-security environments, has highlighted the limitations of traditional facial recognition systems. Models based on fixed similarity thresholds or purely convolutional architectures tend to fail when a large part of the face is occluded, creating a significant gap between laboratory results and real-world deployment. In this context, attention techniques guided by periocular landmarks and hybrid transformers are marking a new paradigm. A recent proposal, the PLGSA-Transformer, addresses the problem by combining heat maps generated from MediaPipe landmarks on the eye, eyebrow, and forehead region, fusing them with features extracted via EfficientNetB3 through a learnable residual gate. This mechanism directs the model's attention toward the most discriminative visible areas. Additionally, it incorporates a hybrid CNN-Transformer branch that converts feature maps into tokens processed by a multi-head self-attention layer, enabling the modeling of dependencies between distant regions. Finally, an occlusion-adaptive similarity threshold dynamically adjusts the matching criterion based on the estimated mask severity, achieving 97.22% accuracy in pair verification with an AUC of 1.0.
For companies that need to integrate robust biometric capabilities into their environments, the development of custom applications that incorporate adaptive artificial intelligence is essential. At Q2BSTUDIO, we work on creating custom software that combines computer vision, attention models, and hybrid architectures to solve real-world problems of access control, authentication, and security. Additionally, we support our clients in deploying these solutions on AI for businesses, leveraging the scalability of AWS and Azure cloud services, and ensuring the cybersecurity of processed biometric data. The integration of AI agents for continuous monitoring and the application of business intelligence services such as Power BI to visualize performance metrics complete a robust technological ecosystem.
The PLGSA-Transformer approach demonstrates that encoding periocular geometry into attention, along with long-range dependency modeling via transformers and adaptive thresholds, offers a scalable and accurate solution for facial recognition with masks. These advances not only close the gap between laboratory and production but also open the door to more inclusive and resilient identification systems. Companies seeking to implement similar technologies can benefit from a comprehensive strategy that combines custom software development, artificial intelligence, and cloud computing, areas in which Q2BSTUDIO brings proven expertise.
