Real-Time Face Mask Compliance Detection
High-Throughput Object Localization with YOLOv5 Architecture
Key contributions
- Trained custom YOLOv5 weights with mosaic augmentations and anchor box tuning.
- Handles partial facial occlusions, varied angles, and diverse illumination levels.
- Structured for edge CCTV camera deployment and live video stream inference.
Applied Machine Learning Engineering Specification
Structured technical analysis conforming to production machine learning standards.
Problem Formulation
Manual verification of public health compliance in crowded facilities is non-scalable, labor-intensive, and prone to human inspection fatigue.
Data Ingestion & Feature Space
Annotated multi-class image dataset with precise bounding box coordinates categorized into three states: masked, unmasked, and incorrectly worn masks.
Preprocessing & Methodology
Engineered robust preprocessing pipeline leveraging mosaic augmentation, color jittering, spatial scaling, and normalized bounding box coordinate conversions.
Model Architecture & Training
Fine-tuned YOLOv5 architecture featuring CSPDarknet53 backbone and Path Aggregation Network (PANet) neck for multi-scale feature localization.
Validation & Evaluation Protocol
Validated against mean Average Precision (mAP@0.5 and mAP@0.5:0.95), inference latency (milliseconds per frame), and precision-recall trade-offs.
Empirical Results & Discoveries
Demonstrated accurate multi-target localization in high-density frames with minimal latency, distinguishing subtle mask placement variations.
Production & Business Impact
Enables contactless, automated compliance telemetry across transit hubs and commercial facilities without requiring on-site personnel.
Tech stack
Want to see the source code?
Access the complete Jupyter notebook, preprocessing routines, and evaluation logs.