Efficient Production of Scene-specific OCR Models Using an AI Platform

Introduction: The article follows up on a previous post about the Laiye OCR testing system, describing how the core OCR capabilities have been optimized on an internal AI platform through data‑driven methods.

Scenario description: Certain OCR use‑cases such as seals, electricity‑meter readings, license plates, tire model numbers, and vehicle VINs have a single visually distinct text region with varying font size, color, or background, making detection easy but recognition challenging.

Model and service production: The AI platform enables rapid creation of OCR models for such scenarios by (1) uploading and annotating data using a customized CVAT tool, (2) training models with YOLOv5 for detection and TrOCR for recognition, (3) publishing the TensorFlow Serving model via a one‑click process, (4) evaluating performance through the platform’s testing system, and (5) exporting a CPU‑compatible Docker image for deployment.

Training pipeline: Data management uses a hierarchical directory with structured tags; training experiments are orchestrated with Kubeflow, leveraging heterogeneous hardware scheduling to run CPU‑only preprocessing and GPU‑accelerated training, and the entire pipeline can produce a new model in under five minutes.

Cost estimation: By employing over 20 image‑augmentation strategies and a 600 million‑image pre‑trained TrOCR model, the system achieves high accuracy with roughly 500‑540 occurrences per character, e.g., 10 000 license‑plate images reach 98 % F1 in about three hours of GPU time.

Conclusion: A data‑centric, AI‑platform approach democratizes OCR model development, allowing business users to obtain effective models simply by collecting and labeling data, while the platform supplies pre‑trained models and automation to keep costs low.

References: Links to the previous blog post, YOLOv5 repository, TrOCR paper, and CVAT repository.