Skip to content

0002-paddleocr-selection.md

Latest commit

 

History

History
45 lines (31 loc) · 1.89 KB

0002-paddleocr-selection.md

File metadata and controls

45 lines (31 loc) · 1.89 KB

2. PaddleOCR for Text Extraction

Date: 2025-11-29

Status

Accepted

Context

The platform requires an Optical Character Recognition (OCR) engine to extract text from uploaded documents (images and PDFs). Requirements:

  1. Accuracy: High accuracy for English and Indian languages (Hindi, Tamil, etc.).
  2. Open Source: Must be open-source and free to use.
  3. Local Execution: Must run locally on the server (no external API calls for data privacy).
  4. Layout Analysis: Ability to detect text blocks and layout is preferred.

Decision

We selected PaddleOCR (specifically the en_PP-OCRv3 and multilingual models) as the primary OCR engine.

Alternatives Considered

1. Tesseract OCR

  • Pros: Standard open-source choice, widely available.
  • Cons: Lower accuracy on complex layouts, requires separate installation of binary (tesseract-ocr), harder to configure for deep learning based layout analysis.

2. EasyOCR

  • Pros: PyTorch-based, easy to install, supports many languages.
  • Cons: Slower than PaddleOCR on CPU, slightly lower accuracy for English scene text compared to PP-OCRv3.

3. Google Cloud Vision / Azure Computer Vision

  • Pros: Best-in-class accuracy.
  • Cons: Paid services, data leaves the premise (privacy concern), requires internet connectivity.

Consequences

Positive

  • High Performance: PaddleOCR offers state-of-the-art accuracy for lightweight models.
  • Language Support: Excellent support for 80+ languages including Indian languages.
  • Deployment: Can be installed via pip (though paddlepaddle dependency can be tricky).

Negative

  • Dependency Size: Requires paddlepaddle which is a large library.
  • Output Format Changes: As experienced during development, PaddleOCR's API output format can vary between versions (list of lists vs list of dicts), requiring robust parsing logic.