Deploy a 24/7 Document Recognition Toolbox with the PaddleOCR Image on the Cloud

PaddleOCR is an open‑source, end‑to‑end OCR and document intelligence engine from Baidu that transforms PDFs or images into structured JSON or Markdown, covering text detection, recognition, layout parsing, and key‑information extraction.

The project has attracted 77,000 GitHub stars and is deeply integrated with popular RAG projects such as RAGFlow and MinerU, making it a common component for building retrieval‑augmented generation and intelligent agents.

PaddleOCR reliably handles real‑world conditions like curved text, scanned documents, screen captures, complex lighting, and skew, outputting fine‑grained coordinates. Its PP‑StructureV3 layout analyzer supports tables (including nested tables), formulas, seals, and charts, providing detailed positional data.

The engine supports over 100 languages and a wide range of document types (Word, Excel, PPT, etc.), offering additional export options such as DOCX. A browser‑side inference SDK, PaddleOCR.js, enables front‑end OCR directly in the browser.

Typical use cases include building intelligent document/RAG pipelines (batch converting PDFs to Markdown/JSON for knowledge bases), automating invoice, receipt, and ID processing for finance or compliance, digitizing forms and tables for downstream analysis, and industrial part or label recognition for manufacturing traceability.

算网’s GPU platform provides a PaddleOCR Docker image that can be deployed as a 24‑hour online OCR service. Users open the platform website, locate the community image, select the desired PaddleOCR version, and confirm the rental.

After logging in with a phone‑based verification code, users choose a GPU instance, start the container, and follow the setup steps:

# Activate Cambricon Neuware driver
export LD_LIBRARY_PATH=/usr/local/neuware/lib64:$LD_LIBRARY_PATH
# Enter project directory
cd /mnt/zk2035044644095582210/magicmind_cloud/buildin/cv/other/paddleocr
# Set dataset path and source environment variables
export ICDAR2015_DATASETS_PATH=/mnt/zk2035044644095582210/temp_data
source env.sh

Next, define the MagicMind model directory and specific model files:

export MM_MODEL_PATH=/mnt/zk2035044644095582210/magicmind_cloud/buildin/cv/other/paddleocr/data/models
export DET_MM_FILE=${MM_MODEL_PATH}/det_model.model
export REC_MM_FILE=${MM_MODEL_PATH}/rec_model.model
export CLS_MM_FILE=${MM_MODEL_PATH}/cls_model.model

Run the inference demo, which automatically reads images from ./doc/imgs and executes on an MLU370 accelerator:

cd infer_python
bash run_demo.sh ${DET_MM_FILE} ${REC_MM_FILE} ${CLS_MM_FILE}

To process custom images, two methods are provided:

Batch mode: place images in /mnt/zk2035044644095582210/my_images and modify the command to point to that directory.

Single‑image mode: specify the absolute path with the --image_dir argument.

After inference, the console prints recognized text, coordinates, and confidence scores. Visual results with red detection boxes are saved under infer_python/system_infer_results/.

Users are encouraged to try the setup and explore PaddleOCR’s capabilities.