Automating Dark Spot Defects Detection in Optical Fiber Bundles Using Machine Learning

Abstract The Optical Fiber Bundle (OFB) consists of multiple optical fibers, each carrying part of an image, with the entire bundle transmitting the image from one location to another. OFBs are highly flexible and lightweight, making them ideal for applications in medical imaging, industrial inspection, and other fields. To ensure performance, OFBs undergo rigorous quality control, with dark spot detection being a significant challenge. Dark spots can be classified as break spots (non-transmitting) or gray spots (partially transmitting), which are difficult to distinguish with the naked eye. This paper proposes an automated method for dark spot detection using machine learning to improve detection efficiency and accuracy. A dark spot database for OFBs was created, and transfer learning with the YOLO object detection model, along with hyperparameter tuning, resulted in an effective model for identifying dark spots. The model achieved around 90% accuracy for detecting images with fewer dark spots, but accuracy dropped to around 85% for images with higher dark spot densities. It introduces a new approach for replacing manual inspection in future OFB production and contributes to the automation and intelligent development of OFB manufacturing processes.