Optical Character Recognition (OCR) remains a fundamental technology, bridging the gap between images and structured text. We’re excited to introduce PP-OCRv6, the latest generation of PaddleOCR’s universal OCR model family, now available on Hugging Face. This release is engineered to tackle the complexities of real-world text extraction, from cluttered documents to multilingual screenshots and industrial labels.
PP-OCRv6 is not just an update; it’s a significant leap forward in performance and versatility. It’s designed to provide accurate, structured text outputs even with compact model sizes and flexible deployment options. For those curious about the ongoing relevance of specialized OCR models in the era of large Vision-Language Models (VLMs), we invite you to explore our previous discussion on PP-OCRv5 and its specialized approach to OCR.
Unpacking PP-OCRv6: Performance and Scalability
The PP-OCRv6 family scales impressively, offering three distinct tiers: tiny (1.5M parameters), small, and medium (up to 34.5M parameters). This design ensures there’s a model suited for every application, balancing accuracy with computational efficiency. The small and medium tiers boast extensive multilingual support, covering 50 languages, including key ones like Simplified Chinese, Traditional Chinese, English, and Japanese, alongside 46 other Latin-script languages.
Performance benchmarks showcase PP-OCRv6’s superiority. On PaddleOCR’s internal multi-scenario OCR benchmarks, the PP-OCRv6_medium model achieves an 86.2% detection Hmean and 83.2% recognition accuracy. This represents a substantial improvement over its predecessor, PP-OCRv5_server, boosting text detection by +4.6 percentage points and text recognition by +5.1 percentage points. You can even try PP-OCRv6 online instantly with the PP-OCRv6 Online Demo.
Innovations Under the Hood
PP-OCRv6 introduces a suite of architectural, training, and data enhancements across both text detection and recognition stages. The core objective was to elevate OCR accuracy while maintaining model sizes that are practical for diverse deployment scenarios. Let’s dive into the key innovations:
- Three Model Tiers: A tailored approach offering tiny, small, and medium models to match varying accuracy and computational requirements.
- PPLCNetV4 Backbone: This unified backbone for both detection and recognition ensures consistency across the entire model family. Developers benefit from a cohesive architecture that underpins all tiers.
- RepLKFPN for Text Detection: Recognizing that accurate detection is crucial, PP-OCRv6 features RepLKFPN, a lightweight large-kernel feature pyramid network. This innovation efficiently handles multi-scale text, making it robust against challenging real-world inputs like small, dense, or rotated text.
- EncoderWithLightSVTR for Recognition: For the recognition phase, PP-OCRv6 employs EncoderWithLightSVTR. This module masterfully combines local context with global attention, significantly enhancing recognition quality on challenging text crops across multilingual, screen, and industrial contexts.
- Unified Multilingual OCR: A standout feature, the medium and small tiers now support 50 languages within a single model family. This dramatically simplifies multilingual OCR workflows, reducing the need for managing multiple specialized models.
Seamless Integration and Flexible Deployment
Getting started with PP-OCRv6 is straightforward, and its flexible architecture supports multiple inference backends. You can quickly integrate PaddleOCR into your projects by installing it via pip:
pip install paddleocrPaddleOCR 3.7 offers a unified inference-engine interface. While Paddle Inference is the default, you can effortlessly switch to other popular backends:
-
Using Paddle Inference (Default):
from paddleocr import PaddleOCR ocr = PaddleOCR( use_doc_orientation_classify=False, use_doc_unwarping=False, use_textline_orientation=False ) result = ocr.predict("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png") for res in result: res.print() res.save_to_img("output") res.save_to_json("output")The results can be saved as visual images or structured JSON output, ready for downstream systems like document parsing, data extraction, RAG pipelines, or analytics.
-
Hugging Face Transformers Backend: For users familiar with Hugging Face, PP-OCRv6 can run seamlessly with a Transformers backend:
from paddleocr import PaddleOCR ocr = PaddleOCR( use_doc_orientation_classify=False, use_doc_unwarping=False, use_textline_orientation=False, engine="transformers" ) result = ocr.predict("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png")This integration streamlines workflows for those operating within the Hugging Face ecosystem. More details are available in the PaddleOCR: Running OCR and Document Parsing Tasks with a Transformers Backend blog post.
-
ONNX Runtime Backend: For environments optimized for ONNX Runtime, PP-OCRv6 also offers ONNX variants within its collection:
from paddleocr import PaddleOCR ocr = PaddleOCR( use_doc_orientation_classify=False, use_doc_unwarping=False, use_textline_orientation=False, engine="onnxruntime" ) result = ocr.predict("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png")These diverse backend options ensure that PP-OCRv6 is accessible and performant across a wide spectrum of runtime environments, all while leveraging the same powerful OCR model family from the Hugging Face Hub.
Conclusion
PP-OCRv6 represents a significant stride in lightweight, multilingual OCR, offering a robust and adaptable solution for real-world text detection and recognition. With model tiers ranging from 1.5M to 34.5M parameters and support for up to 50 languages, it delivers improved accuracy over previous versions.
The models are available in multiple formats on the Hugging Face Hub, including safetensors, Paddle inference models, and ONNX models. Whether you prefer to evaluate it via the online demo, explore the model assets, or integrate it using your preferred inference backend, PP-OCRv6 offers unparalleled flexibility for your OCR workflows.
For further exploration, consult the Transformers Backend Blog, the PaddleOCR Documentation on PP-OCRv6, or the PaddleOCR Official Website. Embrace the future of efficient and accurate OCR with PP-OCRv6!
Source: Hugging Face Blog
