Face mask detection has gained significant attention over the past decade, particularly during the COVID-19 pandemic, where automated monitoring became essential for public health compliance. While early approaches relied on traditional computer vision techniques like Haar cascades and HOG-SVM, recent advancements in deep learning—especially CNNs and transformer-based models—have significantly improved detection accuracy. However, real-world challenges such as varying lighting, occlusions, and diverse mask types continue to hinder robustness. This paper presents an optimized data augmentation framework to enhance mask detection under real-world conditions. Unlike prior works that focus on generic augmentations, we introduce three novel strategies: (1) adaptive geometric transformations that account for facial structure, (2) dynamic photometric adjustments for lighting invariance, and (3) synthetic occlusion generation to improve partial-mask recognition. Our approach builds on YOLOv8, incorporating a modified attention-based neck for small-mask detection. Evaluated on the Kaggle Face Mask Detection dataset, our method achieves 88.7% mAP@0.5, outperforming baseline models by 12.6%. Notably, it shows a 15.3% improvement in occluded scenarios and 10.8% better accuracy in low-light conditions compared to state-of-the-art methods (2020–2023). Despite the computational overhead of advanced augmentations, the system maintains real-time performance (31 FPS on an NVIDIA Jetson Xavier), making it viable for edge deployment. This work bridges a critical gap between laboratory performance and real-world applicability, addressing limitations in prior studies that either overemphasized accuracy on curated datasets or ignored runtime constraints. Future extensions could explore 3D-aware augmentations and federated learning for privacy-sensitive environments.

Keywords: Adaptive Data Augmentation, Occlusion-Robust Detection, YOLOv8 Optimization, Edge-Deployable Vision, Lighting-Invariant Recognition, Pandemic Preparedness