Bounding Box-Guided Diffusion for Synthesizing Industrial Images and Segmentation Maps.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Synthetic dataset generation for industrial applications, especially for defect segmentation, is an underexplored area.
• Acquiring accurate labels for industrial defect segmentation is costly and time-consuming.
• Existing methods lack sufficient defect consistency and spatial accuracy.

Purpose of the Study:

• To propose a novel diffusion-based pipeline for generating high-fidelity industrial datasets with minimal supervision.
• To address the challenges of cost and time associated with acquiring accurate industrial defect data.
• To improve defect consistency and spatial accuracy in synthetic data generation.

Main Methods:

• Utilizing a diffusion model conditioned on enriched bounding-box representations.
• Generating precise segmentation masks for realistic and accurately localized defect synthesis.
• Introducing two novel quantitative metrics for evaluating synthetic data effectiveness.
• Assessing the impact of synthetic data on a downstream segmentation task.

Main Results:

• The proposed diffusion-based pipeline generates high-fidelity industrial datasets.
• The method improves defect consistency and spatial accuracy compared to existing approaches.
• The synthetic data effectively bridges the gap between artificial and real-world industrial data.
• Downstream segmentation models trained on synthetic data show improved reliability and cost-efficiency.

Conclusions:

• Diffusion-based synthesis is a viable solution for generating industrial defect datasets.
• The proposed method offers a more reliable and cost-efficient approach to training segmentation models.
• This work advances the field of synthetic data generation for industrial computer vision applications.