Synthetic data generation framework for quality control automation in gravure printing
By Korota Arsène Coulibaly, Mohamed Hamlich, Khalid Hmali, Andrea Trombin
"A pipeline generates high-fidelity synthetic printing defect images and annotations; training RFDETR on 7533 synthetic images yields 80.9% mAP on real samples, eliminating manual data collection."
Abstract
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 in rotogravure printing. Deep learning models give prospects for automation. However, training robust deep learning models, such as YOLO or Vision Transformers, is heavily hindered by the extreme scarcity of real-world industrial defects images. To overcome this limitation, this paper introduces a novel synthetic data generation framework tailored for rotogravure printing quality control. The proposed pipeline automatically generates high-fidelity images of specific printing defects (creases, streaks, misregistration, etc.) and outputs corresponding bounding boxes and annotations. To validate the framework, a synthetic dataset of 7533 images was generated and used to train the state-of-the-art object-detection model RFDETR. Experimental results demonstrate that the model trained on our synthetic data achieves a Mean Average Precision (mAP) of 80.9\% on real industrial testing samples. This framework provides a zero-cost, rapid-deployment solution for automating defect inspection in printing lines without requiring massive manual data collection.
Technical Analysis & Implementation
Overview§
This paper addresses the scarcity of labeled defect images in rotogravure printing by introducing a synthetic data generation framework. The pipeline automatically creates realistic defect images (e.g., creases, streaks, misregistration) with corresponding bounding boxes. The synthetic dataset is used to train the object detection model RFDETR, achieving strong generalization to real-world samples without any real training data.
Synthetic Data Generation§
The framework takes defect-free template images (e.g., scanned prints) and overlays synthetic defects. Defects are generated by: (1) modeling physical attributes (e.g., Gaussian noise for streaks, edge distortions for misregistration), (2) blending via alpha compositing with realistic opacity, and (3) applying post-processing (jitter, blur) to match real camera artifacts. Bounding boxes are automatically computed from defect masks. 7,533 images were generated with 5 defect classes.
Model Training and Results§
RFDETR (a variant of DETR with efficient attention) was trained on the synthetic dataset using the standard DETR loss: \[ \mathcal{L} = \lambda_{\text{cls}} \mathcal{L}_{\text{cls}} + \lambda_{\text{L1}} \mathcal{L}_{\text{L1}} + \lambda_{\text{giou}} \mathcal{L}_{\text{giou}} \] where $\mathcal{L}_{\text{cls}}$ is cross-entropy for classification, $\mathcal{L}_{\text{L1}}$ and $\mathcal{L}_{\text{giou}}$ are the box regression losses. Hungarian matching aligns predictions with ground truth. The model achieved 80.9% mAP on a real test set, demonstrating the effectiveness of the synthetic augmentation.
Code Snippet (Illustrative Training Loop)§
import torch
from RFDETR import RFDETR
model = RFDETR(num_classes=5, num_queries=100)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
for images, targets in synthetic_dataloader:
outputs = model(images)
loss_dict = criterion(outputs, targets)
loss = sum(loss_dict.values())
optimizer.zero_grad()
loss.backward()
optimizer.step()The criterion uses the Hungarian algorithm from torchvision.detection.
Key Insights§
- Synthetic data at zero cost can replace manual collection for industrial defect detection.
- Domain gap is minimized by physically accurate defect rendering and camera post-processing simulation.
- RFDETR's transformer-based architecture benefits from synthetic data due to its global context understanding.
- This approach is readily transferable to other printing processes.