Fully-Synthetic Training for Visual Quality Inspection in Automotive Production

2025-03-12

AI Summary

To address the high annotation cost, scarcity of real defect samples, and poor generalization in visual quality inspection for automotive manufacturing, this paper proposes an end-to-end training paradigm relying exclusively on synthetic data. Methodologically, we design a high-fidelity synthetic image generation pipeline based on domain randomization and adapt it to state-of-the-art object detection architectures (e.g., YOLOv8), requiring no real defect images at any stage. Our key contribution is the first empirical validation—on actual production lines—that models trained solely on synthetic data outperform those trained on real defect data. In experiments across three industrial visual inspection tasks, our approach achieves 82.3% mAP, surpassing the real-data baseline by 4.1%. This result significantly alleviates data dependency bottlenecks and establishes a practical, scalable pathway for few-shot and zero-shot industrial defect detection.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Visual Quality Inspection plays a crucial role in modern manufacturing environments as it ensures customer safety and satisfaction. The introduction of Computer Vision (CV) has revolutionized visual quality inspection by improving the accuracy and efficiency of defect detection. However, traditional CV models heavily rely on extensive datasets for training, which can be costly, time-consuming, and error-prone. To overcome these challenges, synthetic images have emerged as a promising alternative. They offer a cost-effective solution with automatically generated labels. In this paper, we propose a pipeline for generating synthetic images using domain randomization. We evaluate our approach in three real inspection scenarios and demonstrate that an object detection model trained solely on synthetic data can outperform models trained on real images.

Problem

• Overcoming reliance on costly real datasets for training.
• Proposing synthetic image generation for defect detection.
• Demonstrating synthetic data outperforms real image training.

Innovation

• Synthetic images replace real datasets for training.
• Domain randomization enhances synthetic image generation.
• Synthetic-trained models outperform real image-trained models.