Overview of OCR Technology and Its Applications on Tencent Cloud

OCR (Optical Character Recognition) enables machines to read handwritten or printed text. Handwritten text is highly variable, while printed text is relatively simpler, yet both present challenges that OCR services must handle.

The talk, originally delivered by senior researcher Ji Yongnan of Tencent Cloud’s Big Data AI Product Center, covers the evolution of OCR, its current cloud services, and practical use cases.

Historical background: OCR dates back to the 1960s–70s for postal code recognition, achieving 92‑93% accuracy. The MNIST dataset (2013) originated from these early applications. Google Drive offered free OCR in 2015, and Tencent Cloud launched free OCR access on May 23, 2023, allowing mobile and any devices to upload images for analysis.

Two dimensions to describe OCR applications are presented:

Format dimension – table‑structured OCR vs. general OCR. Table OCR deals with fixed layouts, while general OCR extracts text from arbitrary images.

Script dimension – printed vs. handwritten. Handwritten text is more difficult, especially when fonts blend printed and cursive styles.

Typical OCR workflow: image acquisition → layout analysis → text detection → character segmentation → recognition (often using CNN for feature extraction and RNN for sequence modeling) → post‑processing with language/semantic rules. Modern approaches favor end‑to‑end deep‑learning models (CNN+RNN, attention mechanisms) over traditional multi‑stage pipelines.

Key performance figures from Tencent Cloud:

Millisecond‑level latency on GPU, slightly higher on CPU.

Handwritten digit recognition accuracy >90%.

Single‑character recognition within 15 ms, complex Chinese characters >80% accuracy.

Service portfolio on Tencent Cloud includes:

General printed‑text OCR (returns text and bounding boxes).

Specialized document OCR: ID cards, driver’s licenses, vehicle plates, bank cards, business cards, invoices, medical reports, logistics waybills.

Challenges highlighted:

Image quality issues: blur, skew, low resolution, lighting variations, reflective surfaces on documents.

Diverse fonts and minority languages with limited training data.

Layout complexity in multi‑format documents (e.g., hospital reports from different hospitals).

Case studies:

ID card recognition – widely used for hotel check‑in, high‑speed rail, and security verification.

Business card recognition – semi‑structured with variable field positions.

License‑plate recognition – applied in parking and traffic management, facing challenges of low‑resolution and illumination.

Bank check and logistics waybill recognition – first real‑world deployment of handwritten OCR, achieving >91% accuracy.

Medical insurance underwriting – OCR extracts structured data from heterogeneous medical documents, combined with expert knowledge and machine‑learning models.

Future directions emphasize expanding OCR to more scenarios, enriching error‑correction libraries, and integrating domain‑specific knowledge to improve robustness.

Q&A segment addresses differences between OCR for exam papers vs. logistics waybills, handling large PDF batches, and incorporating prior knowledge into models.