TL;DR

PaddleOCR v3.7.0 introduces PP-OCRv6, a new model offering significantly improved accuracy, unified 50-language support, and faster inference speeds across various hardware (CPU, Apple M4, GPU).

New

• PP-OCRv6 Model: Delivers substantial accuracy gains over the previous version, rivaling large vision-language models with a fraction of the parameters.
• Multi-Language Support: A single model now supports 50 languages, simplifying deployment and eliminating the need for model switching.
• Optimized Performance: Inference is accelerated on CPU (OpenVINO), Apple M4, and A100 GPUs.

Fixes Worth Knowing

• Improved recognition of challenging text types like digital displays, dot-matrix characters, and industrial text.

TL;DR

PaddleOCR now includes a new Chinese recognition model (ch_doc) optimized for documents with a large character set, improving accuracy on complex texts.

Breaking

None.

New

• Chinese Document Recognition: Added ch_doc language model for improved accuracy on document images with over 15,000 characters.
• Environment Variable for Models: You can now specify the base directory for PaddleOCR models using the PADDLE_OCR_BASE_DIR environment variable.

Fixes Worth Knowing

• Fixed NaN issues in PP-OCRv4, improving stability.
• Resolved inference errors with KIE (Key Information Extraction) mode in ONNX format.
• Corrected issues with LaTeXOCR inference and training.
• Fixed a bug causing errors when exporting images without text to docx format.

Before You Upgrade

If you were relying on specific model download locations, consider setting the PADDLE_OCR_BASE_DIR environment variable to ensure consistent behavior.

TL;DR

PaddleOCR now offers a 3.5MB ultra-lightweight OCR (Optical Character Recognition) system, enabling deployment on a wider range of devices including mobile and embedded systems.

New

• Ultra-lightweight model: Deploy OCR on server, mobile, embedded, and IoT devices.

Fixes Worth Knowing

None.

Before You Upgrade

None.