Synthetic data generation framework for quality control automation in rotogravure printing

Quality control in rotogravure printing is traditionally a manual, subjective, and labor-intensive process. While deep learning offers a path toward automated inspection, the primary bottleneck is the scarcity of high-quality, pixel-annotated training data. Real-world defects are rare, and manually labeling thousands of images is often prohibitively expensive and time-consuming for industrial projects. This paper addresses this challenge by proposing a synthetic data generation framework that produces realistic, annotated defect images at zero cost.

The proposed framework operates by taking clean, defect-free images and programmatically injecting common printing anomalies—specifically fisheyes, streaks, misregistration, and creases. Each defect type is modeled using physical and geometric principles:

• Fisheyes: Modeled as circular contamination spots that alter pixel intensity based on distance from the center.
• Streaks: Modeled as vertical photometric alterations using Perlin noise to simulate organic, non-uniform ink distribution.
• Misregistration: Modeled by simulating mechanical shifts and ink-spreading effects across different color cylinders.
• Creases: Modeled by combining geometric distortion (pinching) with photometric effects like shadows and specular reflections.

Because the framework generates the defect and the corresponding segmentation mask simultaneously, it provides perfect, pixel-level ground truth labels without human intervention.

To validate the approach, the authors generated a synthetic dataset of 7,533 images. They trained the state-of-the-art RFDETR object-detection model on this synthetic data and evaluated its performance on real industrial testing samples. The model achieved a Mean Average Precision (mAP) of 80.9%, demonstrating that synthetic data can effectively bridge the gap between limited real-world samples and the high data requirements of modern deep learning architectures.