On Vision Transformer Explainability for Personal Protective Equipment Detection: A Qualitative and Quantitative...

Area of Science:

• Industrial Safety
• Artificial Intelligence
• Machine Learning

Background:

• Worker safety in industrial settings relies on correct Personal Protective Equipment (PPE) usage.
• Deep Learning (DL) offers potential for monitoring PPE compliance.
• Federated Machine Learning (FML) enables privacy-preserving and explainable AI for sensitive data.

Purpose of the Study:

• To evaluate the robustness and consistency of explainability algorithms in identifying crucial areas for PPE classification.
• To assess the effectiveness of similarity indices in validating explainability models within an FML framework.
• To analyze the performance of explainability techniques in real-world industrial safety monitoring.

Main Methods:

• Utilized a dataset of 1600 real-world images of workers with and without various PPE (helmets, reflective vests).
• Employed Federated Machine Learning (FML) for privacy-preserving analysis.
• Applied similarity indices such as Structural Similarity Index Measure (SSIM), Visual Information Fidelity (VIF), and Similarity-based Consistency Check (SCC) to evaluate explainability.

Main Results:

• High mean similarity index values were observed, indicating robust explainability.
• Intra-client study yielded mean values of 0.99 (SSIM), 0.96 (VIF), and 0.96 (SCC).
• Inter-client analysis showed mean values of 0.96 (SSIM), 0.91 (VIF), and 0.71 (SCC), demonstrating consistent performance across different data distributions.

Conclusions:

• The study confirms the robustness and consistency of explainability algorithms for PPE classification using FML.
• The employed similarity indices effectively validate the reliability of AI models in identifying critical safety elements.
• FML-powered explainable AI shows significant promise for enhancing industrial safety monitoring and compliance.