Introduction
Scene Text Recognition (STR)—the ability of AI to read text in the wild—traditionally relies on massive, centralized datasets. However, real-world data is often "trapped" on local devices due to strict privacy regulations like GDPR. While Federated Learning (FL) offers a way to train models on decentralized data without moving the images themselves, applying it to STR is difficult due to high communication overhead and potential privacy leakage from gradient analysis.
Methodology: Efficiency and Security via Hashing
The authors propose FedOCR, a horizontal federated learning framework. It addresses the dual hurdles of data privacy and communication costs by combining lightweight neural architectures with a clever hashing technique.
The framework replaces standard heavyweight backbones (like ResNet) with ShuffleNetV2, reducing the initial parameter count for edge devices. To further optimize communication and security, FedOCR employs a Hashing Trick. Instead of transmitting full weight matrices, the model trains a small "real" weight vector. A hashing function and random seed map these values into a larger "virtual" weight matrix. This provides compression (reducing update sizes) and security (obfuscating the model updates from the central server).
The FedOCR process follows a four-step cycle:
- Global parameters downloading: Clients receive the current global model.
- Local training and update: Clients train on their private, local data.
- Local parameters uploading: Clients send optimized, compressed updates back to the server.
- Aggregation: The server aggregates updates to form a new global model.
Experimental Results
The study utilized a comprehensive suite of standard STR datasets (Synth90k, SynthText) and real-world benchmarks (IC03, IC13, IIIT5K, etc.). Key findings include:
- Parameter Reduction: The base FedOCR is 36% smaller than the ResNet-based ASTER-FL. With the hashing trick at a 1/8 ratio, the model size drops from 80.52 MB to just 7.05 MB (a 91% reduction).
- Competitive Accuracy: FedOCR-Hash$_1$ achieved 89.08% average accuracy, very close to the 91.94% of the much heavier ASTER-FL baseline.
- Communication Speed: FedOCR reached high accuracy levels much earlier than the baseline in terms of total megabytes uploaded. For example, FedOCR-Hash$_{1/4}$ requires only 13.38 MB of transmission per participant per round.
Conclusion and Future Work
FedOCR demonstrates that privacy-preserving STR is feasible without sacrificing significant accuracy or bandwidth. The authors are currently working on integrating domain adaptation techniques to address "Domain Shift"—the performance degradation caused by non-IID data across diverse, decentralized data sources.
Summary updated