PaddleOCR

• The default download source has been changed from BOS to HuggingFace. Users can also change the environment variable PADDLE_PDX_MODEL_SOURCE to BOS to set the model download source back to Baidu Object Storage (BOS).

• Added service invocation examples for six languages—C++, Java, Go, C#, Node.js, and PHP—for pipelines like PP-OCRv5, PP-StructureV3, and PP-ChatOCRv4.

• Improved the layout partition sorting algorithm in the PP-StructureV3 pipeline, enhancing the sorting logic for complex vertical layouts to deliver better results.

• Enhanced model selection logic: when a language is specified but a model version is not, the system will automatically select the latest model version supporting that language.

• Set a default upper limit for MKL-DNN cache size to prevent unlimited growth, while also allowing users to configure cache capacity.

• Updated default configurations for high-performance inference to support Paddle MKL-DNN acceleration and optimized the logic for automatic configuration selection for smarter choices.

• Adjusted the logic for obtaining the default device to consider the actual support for computing devices by the installed Paddle framework, making program behavior more intuitive.

• Added Android example for PP-OCRv5. Details.

• Bug Fixes:

  • Fixed an issue with some CLI parameters in PP-StructureV3 not taking effect.
  • Resolved an issue where export_paddlex_config_to_yaml would not function correctly in certain cases.
  • Corrected the discrepancy between the actual behavior of save_path and its documentation description.
  • Fixed potential multithreading errors when using MKL-DNN in basic service deployment.
  • Corrected channel order errors in image preprocessing for the Latex-OCR model.
  • Fixed channel order errors in saving visualized images within the text recognition module.
  • Resolved channel order errors in visualized table results within PP-StructureV3 pipeline.
  • Fixed an overflow issue in the calculation of overlap_ratio under extremely special circumstances in the PP-StructureV3 pipeline.

• Documentation Improvements:

  • Updated the description of the enable_mkldnn parameter in the documentation to accurately reflect the program's actual behavior.
  • Fixed errors in the documentation regarding the lang and ocr_version parameters.
  • Added instructions for exporting pipeline configuration files via CLI.
  • Fixed missing columns in the performance data table for PP-OCRv5.
  • Refined benchmark metrics for PP-StructureV3 across different configurations.

• Others:

  • Relaxed version restrictions on dependencies like numpy and pandas, restoring support for Python 3.12.

2025.06.05: PaddleOCR 3.0.1 Released, includes:

• Optimisation of certain models and model configurations:
  • Updated the default model configuration for PP-OCRv5, changing both detection and recognition from mobile to server models. To improve default performance in most scenarios, the parameter limit_side_len in the configuration has been changed from 736 to 64.
  • Added a new text line orientation classification model PP-LCNet_x1_0_textline_ori with an accuracy of 99.42%. The default text line orientation classifier for OCR, PP-StructureV3, and PP-ChatOCRv4 pipelines has been updated to this model.
  • Optimized the text line orientation classification model PP-LCNet_x0_25_textline_ori, improving accuracy by 3.3 percentage points to a current accuracy of 98.85%.
• Optimizations and fixes for some issues in version 3.0.0, details

🔥🔥2025.05.20: Official Release of PaddleOCR v3.0, including:

• PP-OCRv5: High-Accuracy Text Recognition Model for All Scenarios - Instant Text from Images/PDFs.

  1. 🌐 Single-model support for five text types - Seamlessly process Simplified Chinese, Traditional Chinese, Simplified Chinese Pinyin, English and Japanese within a single model.
  2. ✍️ Improved handwriting recognition: Significantly better at complex cursive scripts and non-standard handwriting.
  3. 🎯 13-point accuracy gain over PP-OCRv4, achieving state-of-the-art performance across a variety of real-world scenarios.

• PP-StructureV3: General-Purpose Document Parsing – Unleash SOTA Images/PDFs Parsing for Real-World Scenarios!

  1. 🧮 High-Accuracy multi-scene PDF parsing, leading both open- and closed-source solutions on the OmniDocBench benchmark.
  2. 🧠 Specialized capabilities include seal recognition, chart-to-table conversion, table recognition with nested formulas/images, vertical text document parsing, and complex table structure analysis.

• PP-ChatOCRv4: Intelligent Document Understanding – Extract Key Information, not just text from Images/PDFs.

  1. 🔥 15-point accuracy gain in key-information extraction on PDF/PNG/JPG files over the previous generation.
  2. 💻 Native support for ERNIE 4.5, with compatibility for large-model deployments via PaddleNLP, Ollama, vLLM, and more.
  3. 🤝 Integrated PP-DocBee2, enabling extraction and understanding of printed text, handwriting, seals, tables, charts, and other common elements in complex documents.

History Log

from pathlib import Path
from paddleocr import PPStructureV3

pipeline = PPStructureV3(
    use_doc_orientation_classify=False,
    use_doc_unwarping=False
)

# For Image
output = pipeline.predict(
    input="https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/pp_structure_v3_demo.png",
)

# Visualize the results and save the JSON results
for res in output:
    res.print() 
    res.save_to_json(save_path="output") 
    res.save_to_markdown(save_path="output")           

from paddleocr import PPChatOCRv4Doc

chat_bot_config = {
    "module_name": "chat_bot",
    "model_name": "ernie-3.5-8k",
    "base_url": "https://qianfan.baidubce.com/v2",
    "api_type": "openai",
    "api_key": "api_key",  # your api_key
}

retriever_config = {
    "module_name": "retriever",
    "model_name": "embedding-v1",
    "base_url": "https://qianfan.baidubce.com/v2",
    "api_type": "qianfan",
    "api_key": "api_key",  # your api_key
}

pipeline = PPChatOCRv4Doc(
    use_doc_orientation_classify=False,
    use_doc_unwarping=False
)

visual_predict_res = pipeline.visual_predict(
    input="https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/vehicle_certificate-1.png",
    use_common_ocr=True,
    use_seal_recognition=True,
    use_table_recognition=True,
)

mllm_predict_info = None
use_mllm = False
# If a multimodal large model is used, the local mllm service needs to be started. You can refer to the documentation: https://github.com/PaddlePaddle/PaddleX/blob/release/3.0/docs/pipeline_usage/tutorials/vlm_pipelines/doc_understanding.en.md performs deployment and updates the mllm_chat_bot_config configuration.
if use_mllm:
    mllm_chat_bot_config = {
        "module_name": "chat_bot",
        "model_name": "PP-DocBee",
        "base_url": "http://127.0.0.1:8080/",  # your local mllm service url
        "api_type": "openai",
        "api_key": "api_key",  # your api_key
    }

    mllm_predict_res = pipeline.mllm_pred(
        input="https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/vehicle_certificate-1.png",
        key_list=["驾驶室准乘人数"],
        mllm_chat_bot_config=mllm_chat_bot_config,
    )
    mllm_predict_info = mllm_predict_res["mllm_res"]

visual_info_list = []
for res in visual_predict_res:
    visual_info_list.append(res["visual_info"])
    layout_parsing_result = res["layout_parsing_result"]

vector_info = pipeline.build_vector(
    visual_info_list, flag_save_bytes_vector=True, retriever_config=retriever_config
)
chat_result = pipeline.chat(
    key_list=["驾驶室准乘人数"],
    visual_info=visual_info_list,
    vector_info=vector_info,
    mllm_predict_info=mllm_predict_info,
    chat_bot_config=chat_bot_config,
    retriever_config=retriever_config,
)
print(chat_result)