FedTR combines federated and transfer learning for 95.5% inspection accuracy
New framework achieves centralized-training parity on limited, private industrial data.
FedTR, developed by Vikash Sathiamoorthy and nine co-authors, tackles a critical challenge in industrial visual inspection: performing accurate defect detection (specifically label defects via end-to-end text recognition) when data is both limited and spread across multiple sites under privacy constraints. Traditional federated learning often suffers when local datasets are small and inspection tasks are complex. FedTR solves this by first training a model on a large public dataset, then using federated fine-tuning on distributed, private industrial data.
In experiments on private ink-cartridge datasets, FedTR achieved an end-to-end text recognition word-level accuracy of 95.5% on homogeneous data and 94.2% on heterogeneous data—performance comparable to centralized training. The paper appears in GLSVLSI 2024 proceedings and demonstrates a practical path for deploying privacy-preserving AI in manufacturing without sacrificing accuracy.
- Achieves 95.5% accuracy on homogeneous and 94.2% on heterogeneous private ink-cartridge data.
- Combines transfer learning (pre-train on public data) with federated fine-tuning on limited private data.
- Matches centralized training performance while fully preserving data privacy across industrial sites.
Why It Matters
Enables privacy-preserving, high-accuracy defect detection in manufacturing with limited data—a practical win for Industry 4.0.