Skip to content

MODELS.md

Latest commit

 

History

History
103 lines (81 loc) · 3.3 KB

MODELS.md

File metadata and controls

103 lines (81 loc) · 3.3 KB

Available OCR Models

This document describes all available OCR models in this project.

Model Sources

All models are sourced from:

Model Configurations

1. PaddleOCR v5 Mobile (Default)

  • ID: ppocrv5-mobile
  • Type: Mobile (Fast)
  • Best For: General purpose, balanced speed and accuracy
  • Languages: English, Chinese
  • Components:
    • Detection: ch_PP-OCRv5_det.onnx (4.6MB)
    • Recognition: ch_PP-OCRv5_rec.onnx (16.5MB)
    • Classification: ch_PP-OCRv5_cls.onnx (583KB)

2. PaddleOCR v4 Mobile

  • ID: ppocrv4-mobile
  • Type: Mobile (Fast)
  • Best For: Stable performance, smaller model size
  • Languages: English, Chinese
  • Components:
    • Detection: ch_PP-OCRv4_det.onnx (4.6MB)
    • Recognition: ch_PP-OCRv4_rec.onnx (10.8MB)
    • Classification: ch_PP-OCRv4_cls.onnx (583KB)

3. PaddleOCR v2 Server

  • ID: ppocrv2-server
  • Type: Server (Accurate)
  • Best For: High accuracy detection, slower processing
  • Languages: Chinese (primary)
  • Components:
    • Detection: ch_ppocr_server_v2.0_det.onnx (47MB)
    • Recognition: ch_PP-OCRv4_rec.onnx (10.8MB) - Fallback from v4
    • Classification: ch_ppocr_server_v2.0_cls.onnx (583KB)

4. English Mobile

  • ID: en-mobile
  • Type: Mobile (Very Fast)
  • Best For: English-only text, fastest processing
  • Languages: English
  • Components:
    • Detection: PP-OCRv5_mobile_det.onnx (4.7MB)
    • Recognition: en_PP-OCRv4_mobile_rec.onnx (8.6MB)
    • Classification: ch_PP-OCRv4_cls.onnx (583KB)

Model Architecture

All models follow the PaddleOCR architecture:

  1. Detection Model (DB)

    • Uses Differentiable Binarization (DB) algorithm
    • Outputs probability maps for text regions
    • Input: RGB image (any size, resized to multiple of 32)
    • Output: Binary map of text regions

  2. Recognition Model (CRNN)

    • Uses CRNN + CTC for text recognition
    • Processes cropped text regions
    • Input: Grayscale image (48px height)
    • Output: Character sequence probabilities

  3. Classification Model (CNN)

    • Detects text orientation (0° or 180°)
    • Used to correct upside-down text
    • Input: Text region (48x192px)
    • Output: Orientation probability

Dictionary Files

Each model configuration includes a dictionary file that maps output indices to characters:

  • ppocr_keys_v1.txt - Standard PaddleOCR dictionary (6623 chars)
  • ppocrv5_dict.txt - Extended v5 dictionary
  • en_dict.txt - English-only dictionary (96 chars)

Performance Comparison

Model Detection Speed Recognition Speed Overall Accuracy
v5 Mobile Fast Fast 95%+
v4 Mobile Fast Very Fast 93%+
v2 Server Slow Fast* 97%+
English Mobile Very Fast Very Fast 94%+ (English)

*v2 Server uses v4 recognition model as fallback

Usage in Code

import { getModelConfig } from './core/model-config'

// Get default model (v5 mobile)
const defaultConfig = getModelConfig('ppocrv5-mobile')

// Initialize with specific model
await engine.initialize(defaultConfig.paths, defaultConfig.id)