FiberBobbin-40K

This repository provides a partial release of FiberBobbin-40K, the benchmark dataset accompanying our paper, “A Real-Time Surface Defect Detection Framework for Synthetic Fiber Bobbins.”

Our work addresses the industrial and edge deployment of surface defect detection for synthetic fiber bobbins. We develop a real-time detection framework featuring a lightweight quantization scheme, a novel model design strategy, and an enhanced training pipeline. The framework integrates automated detection-head search, structural recombination, model updating, pruning, and quantization, enabling efficient real-time inference on edge devices.

In addition, we design an integrated posture-adjustment device for synthetic-fiber bobbins and establish FiberBobbin-40K, a dedicated industrial defect detection dataset. To balance research openness with laboratory asset management, we release a representative subset that supports preliminary training, validation, and subsequent research.