Distribution-guided synthetic data generation for data-scarce industrial inspection

Industrial object detection systems frequently operate under severe data scarcity and long-tailed class distributions, particularly in manufacturing environments characterized by small components, structured layouts, and controlled acquisition conditions. Under these constraints, conventional synthetic generation strategies often fail to preserve the spatial and statistical properties required for reliable industrial deployment. This work proposes a distribution-guided synthetic data generation framework for industrial inspection based on constrained copy-paste compositing. Starting from a limited set of annotated images, object instances are extracted and recomposed onto reconstructed backgrounds to generate fully labeled synthetic training samples without additional annotation effort. Unlike conventional copy-paste augmentation, the proposed framework explicitly regulates instance-level class distributions during generation while preserving dataset-derived spatial characteristics, including object-density patterns and realistic occlusion statistics. Object placement is further constrained using empirical overlap distributions computed from the original dataset, enabling the synthesis of structurally consistent industrial scenes. Experimental evaluation on a proprietary industrial dataset demonstrates consistent improvements over standard augmentation, compositional augmentation, and diffusion-based synthetic generation baselines. The proposed method achieves the highest detection accuracy while maintaining lower computational complexity and faster convergence during training. A component-wise analysis further shows that explicit instance-level distribution control is the primary factor driving improvements on underrepresented categories, while image-level augmentation provides complementary gains when integrated into the full pipeline. Additional experiments on an external open-access industrial dataset confirm the robustness and generalization capability of the proposed framework across different inspection scenarios. Overall, the proposed approach provides a lightweight, controllable, and deployment-oriented solution for synthetic data generation in data-scarce industrial object detection systems.

Authors

Publication Details

Journal
The International Journal of Advanced Manufacturing Technology
Published
2026-11-20
DOI
https://doi.org/10.1007/s00170-026-19352-3
Primary Topic
Industrial Vision Systems and Defect Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Distribution-guided synthetic data generation for data-scarce industrial inspection

Cigdem Beyan, Matteo Cristani, Alessandro Venaruzzo
The International Journal of Advanced Manufacturing Technology
Industrial Vision Systems and Defect Detection
article

Distribution-guided synthetic data generation for data-scarce industrial inspection

Cigdem Beyan, Matteo Cristani, Alessandro Venaruzzo
article en

Abstract

Industrial object detection systems frequently operate under severe data scarcity and long-tailed class distributions, particularly in manufacturing environments characterized by small components, structured layouts, and controlled acquisition conditions. Under these constraints, conventional synthetic generation strategies often fail to preserve the spatial and statistical properties required for reliable industrial deployment. This work proposes a distribution-guided synthetic data generation framework for industrial inspection based on constrained copy-paste compositing. Starting from a limited set of annotated images, object instances are extracted and recomposed onto reconstructed backgrounds to generate fully labeled synthetic training samples without additional annotation effort. Unlike conventional copy-paste augmentation, the proposed framework explicitly regulates instance-level class distributions during generation while preserving dataset-derived spatial characteristics, including object-density patterns and realistic occlusion statistics. Object placement is further constrained using empirical overlap distributions computed from the original dataset, enabling the synthesis of structurally consistent industrial scenes. Experimental evaluation on a proprietary industrial dataset demonstrates consistent improvements over standard augmentation, compositional augmentation, and diffusion-based synthetic generation baselines. The proposed method achieves the highest detection accuracy while maintaining lower computational complexity and faster convergence during training. A component-wise analysis further shows that explicit instance-level distribution control is the primary factor driving improvements on underrepresented categories, while image-level augmentation provides complementary gains when integrated into the full pipeline. Additional experiments on an external open-access industrial dataset confirm the robustness and generalization capability of the proposed framework across different inspection scenarios. Overall, the proposed approach provides a lightweight, controllable, and deployment-oriented solution for synthetic data generation in data-scarce industrial object detection systems.

The International Journal of Advanced Manufacturing Technology
Industrial Vision Systems and Defect Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon