Automated Assessment of PPE Compliance with Fast Lightweight Deep Learning Based Computer Vision

This study presents a fast, lightweight computer vision framework that achieves high accuracy in verifying the correct use of goggles and N95 masks. For ergonomics practitioners, this model offers a computer vision tool for enhancing workplace safety compliance, that operates up to five times faster than benchmark models. The proposed algorithm enables automated PPE compliance checks at worksite entrances using consumer-grade computing hardware. Thus, reducing the burden of daily safety checks. This system can directly reduce preventable injuries through ensuring workers are properly protected by PPEs before starting work.

Ensuring the proper usage of Personal Protective Equipment (PPE) is mandatory to minimize workplace injuries. U.S. Department of Labor reported that appropriate PPE usage reduces over 37% of occupational diseases and injuries. Previous research investigated RepVGG and ResNet50 for PPE compliance, but the highest accuracy reported is limited, about 80%. Another challenge of PPE compliance research is the tradeoff between large model size, with high inference times, and achieving satisfactory accuracy.

Our study presents a novel lightweight computer vision framework for automated PPE compliance verification prior to worksite entry. The objective is to achieve 90% of classification accuracy with faster inference time than standard Vision Transformer, while allowing potential deployment on consumer-level computing hardware. Another objective is to identify the incorrect PPE wearing scenario with the spatial relationship awareness between workers and their equipment.

The proposed method constructs a state space model layer for linear computation modeling and a spatial attention module to capture global spatial dependencies with positional encoding of an image. We utilized two public PPE compliance datasets, PPE-CLS and PPE-BQZEL, to evaluate the goggles and N95 mask compliance in manufacturing.

The model achieved 0.91 and 0.94 accuracy on two datasets, respectively. This outperforms a standard Vision Transformer (ViT) which achieved 0.88 and 0.90, while operating up to 5 times faster. The framework's minimal size (2.8 M parameters) and linear inference scaling enable real-time performance without sacrificing accuracy.

This work provides a fast, memory-efficient solution for automated safety supervision, which advances the artificial intelligence and machine learning applications for occupational health.