Generation of synthetic images to expand datasets for computer vision systems used on robotic conveyors

Abstract

The paper presents a comprehensive approach to expanding datasets using synthetic image generation for advanced computer vision systems applied to robotic sorting in automated glass product quality control systems. The proposed method integrates synthetic data generation with deep learning models for defect detection in glass container production. The system combines composite rendering using the Pillow library and automatic annotation generation in COCO format to create photorealistic variations of original objects with controlled variability in lighting and background conditions. The defect generation module integrated into the computer vision system realistically simulates production defects in glass containers, including chips, cracks, microbubbles, scratches, label wear, and shape deformations, significantly enhancing data variability and realism. Experimental validation demonstrated the method’s effectiveness for generating balanced datasets for training robust computer vision models. A cascaded approach using YOLO for rapid region of interest extraction, EfficientDet and RetinaNet for precise defect identification, and Vision Transformer for verification achieves high accuracy in real-time industrial conditions. The system successfully generated 1000 synthetic images in approximately 15 seconds, maintaining 100% annotation accuracy and area deviation of less than 1%. This approach demonstrates high performance and is designed for integration into a comprehensive computer vision system architecture using cascaded models (YOLO, EfficientDet, Vision Transformer) for real-time quality control on production lines, effectively addressing class imbalance and data scarcity problems characteristic of industrial applications.