Enhancing industrial safety AI with synthetic data on Amazon SageMaker AI

This article demonstrates a synthetic data augmentation pipeline for industrial safety AI that generates photo-realistic training images with automated labels to address the scarcity of hazardous scenario data.

• Uses Qwen-Image-Edit-2509 diffusion model on Amazon SageMaker AI to insert synthetic people into real equipment images
• Applies Amazon Rekognition DetectLabels API for automated pseudo-labeling without manual annotation
• Domain-relevant hazardous placement of synthetic people doubled person detection mAP50 performance
• Optimal synthetic data volume (750 images) achieved 160% improvement in person detection mAP50 over baseline
• YOLO11-medium model with synthetic data doubled person recall (17% to 34%) while maintaining edge deployability
• Reduces per-image cost from $3–$5 (manual) to $0.33 (synthetic) and eliminates safety risks of staging dangerous scenarios
• Complete open-source implementation available on GitHub with modular prompt templates for different industrial domains

The pipeline enables safety-critical industrial AI systems to achieve superior detection performance for rare hazardous scenarios without manual data collection or dangerous photography sessions.