Presentation Information
[1I3-GS-10f-05]Establishment of a Visual Inspection Method that is Robust Against Changes in Shooting Conditions
〇SATORU SHIMMURA1 (1. Nagano Prefecture General Industrial Technology Center)
Keywords:
Visual inspection,Anomaly detection
In recent years, AI-based visual inspection has achieved high accuracy on public datasets. However, applying these models to real-world manufacturing remains challenging due to accuracy degradation caused by fluctuating imaging conditions, such as misalignment and lighting changes. Particularly in low-cost conveyor-based systems, these variations are inevitable, rendering conventional methods insufficient.
While zero-shot anomaly detection and 3D Gaussian Splatting have been proposed as solutions, they face limitations: the former often overreacts to specific geometric features, while the latter increases costs and reduces inspection speed by requiring multiple cameras.
To overcome these issues, we propose AprilGAN-FA, a system that achieves high-precision inspection using a cost-effective single camera and a limited number of anomalous samples. By integrating zero-shot detection with few-shot learning, our method ensures robustness against environmental fluctuations. This approach enables the development of a practical, high-speed inspection system while minimizing the time and cost of data collection.
While zero-shot anomaly detection and 3D Gaussian Splatting have been proposed as solutions, they face limitations: the former often overreacts to specific geometric features, while the latter increases costs and reduces inspection speed by requiring multiple cameras.
To overcome these issues, we propose AprilGAN-FA, a system that achieves high-precision inspection using a cost-effective single camera and a limited number of anomalous samples. By integrating zero-shot detection with few-shot learning, our method ensures robustness against environmental fluctuations. This approach enables the development of a practical, high-speed inspection system while minimizing the time and cost of data collection.
Comment
To browse or post comments, you must log in.Log in