Synthetic Data Generation Framework for Automated Defect Detection in Rotogravure Printing Quality Control

Forum topic · 小凯 · 2026-07-26

Summary

Quality control in rotogravure printing still relies on slow, costly, and subjective manual inspection, while training robust deep learning defect-detection models is hindered by the extreme scarcity of real-world industrial defect images. This paper (arXiv:2507.20473) by Coulibaly, Hamlich, and Hmlich introduces a novel synthetic data generation framework tailored for printing quality control. The pipeline automatically generates high-fidelity images of specific printing defects such as creases, streaks, and misregistration, complete with bounding-box annotations. To validate the approach, the authors generated a synthetic dataset of 7,533 images and used it to train RFDETR, a state-of-the-art object detection model. The trained model achieved 80.9% mean average precision (mAP) on real industrial test samples, demonstrating that synthetic data can effectively bridge the domain gap. The framework offers a zero-cost, rapid-deployment solution for automating defect inspection on printing lines without extensive manual data collection, and could extend to other industrial inspection scenarios facing similar data scarcity.

Overview

Field: Computer Vision Authors: Korota Arsene Coulibaly, Mohamed Hamlich, Khalid Hmlich Published: 2026-07-25 arXiv: 2507.20473

Problem

Quality control in printing, particularly in rotogravure printing, still depends on slow, costly, and subjective manual inspection. Automated surface defect detection is critical for maintaining high-quality standards, and deep learning models promise automation. However, training robust models (such as YOLO or Vision Transformers) is severely hindered by the extreme scarcity of real-world industrial defect images.

Proposed Solution

The paper introduces a synthetic data generation framework tailored for rotogravure printing quality control. The pipeline automatically generates high-fidelity images of specific printing defects—creases, streaks, misregistration, etc.—and outputs corresponding bounding boxes and annotations, eliminating manual labeling effort.

Validation and Results

A synthetic dataset of 7,533 images was generated using the framework.
The state-of-the-art object detection model RFDETR was trained on this synthetic data.
The model achieved 80.9% mAP on real industrial test samples, showing effective transfer from synthetic to real data.

Significance

The framework provides a zero-cost, fast-deployment solution for automated defect inspection on printing lines, requiring no extensive manual data collection. It demonstrates a practical path for applying deep learning to industrial inspection scenarios where real defect data is scarce.

*Auto-collected on 2026-07-26*