Model Selection and Language Support

Relevant source files

• .clang_format.hook
• README.md
• deploy/lite/Makefile
• deploy/lite/cls_process.cc
• deploy/lite/cls_process.h
• deploy/lite/config.txt
• deploy/lite/crnn_process.cc
• deploy/lite/crnn_process.h
• deploy/lite/db_post_process.cc
• deploy/lite/db_post_process.h
• deploy/lite/imgs/lite_demo.png
• deploy/lite/ocr_db_crnn.cc
• deploy/lite/prepare.sh
• deploy/lite/readme.md
• docs/index.en.md
• docs/index.md
• docs/quick_start.en.md
• docs/quick_start.md
• docs/version3.x/pipeline_usage/instructions/parallel_inference.en.md
• docs/version3.x/pipeline_usage/instructions/parallel_inference.md
• mkdocs.yml

1. Purpose and Scope

This document guides users through selecting appropriate OCR models based on three criteria: language requirements, OCR version (PP-OCRv5, PP-OCRv4, PP-OCRv3), and deployment target (server vs. mobile). Model selection involves trade-offs between accuracy, inference speed, model size, and language coverage.

Key Selection Dimensions:

• Language: Single-model multilingual (PP-OCRv5), language-specific models (20+ languages), or universal VLM (111+ languages)
• Version: PP-OCRv5 (default, 13% accuracy improvement), PP-OCRv4 (15K+ character support), PP-OCRv3 (legacy)
• Scale: Server models (higher accuracy, 80-180MB) vs. Mobile models (efficiency-optimized, 4-16MB)

Model selection directly impacts pipeline configuration through parameters like text_detection_model_name, text_recognition_model_name, ocr_version, and lang in the PaddleOCR class paddleocr/__init__.py and CLI paddleocr ocr command paddleocr/__main__.py

Sources: docs/version3.x/pipeline_usage/OCR.md1-241 docs/version3.x/pipeline_usage/OCR.en.md1-240 pyproject.toml6-61

2. Model Version Selection

2.1 Version Overview and Model Registry Names

PaddleOCR provides three major versions with distinct model registry names used in text_detection_model_name and text_recognition_model_name parameters:

Title: PP-OCR Version Evolution and Model Names

Version Selection Criteria:

Version	When to Use	Model Registry Names
PP-OCRv5	Default choice; multilingual support (Chinese, English, Traditional Chinese, Japanese); highest accuracy	PP-OCRv5_server_det, PP-OCRv5_mobile_det, PP-OCRv5_server_rec, PP-OCRv5_mobile_rec
PP-OCRv4	Need 15K+ character dictionary (_doc variant); existing deployment compatibility	PP-OCRv4_server_det, PP-OCRv4_mobile_det, PP-OCRv4_server_rec, PP-OCRv4_server_rec_doc, PP-OCRv4_mobile_rec
PP-OCRv3	Legacy system integration only	PP-OCRv3_mobile_rec

Configuration Examples:

Sources: docs/version3.x/pipeline_usage/OCR.md11-240 docs/version3.x/pipeline_usage/OCR.en.md11-239 docs/version3.x/pipeline_usage/OCR.md132-169

2.2 Version Selection Decision Criteria

Sources: README.md67-70 docs/version3.x/pipeline_usage/OCR.md183-240 docs/version3.x/pipeline_usage/OCR.en.md183-239

3. Model Scale Selection: Server vs. Mobile

3.1 Scale Variant Characteristics

PaddleOCR provides two scale variants for most model versions:

Characteristic	Server Models	Mobile Models
Primary Goal	Maximum accuracy	Deployment efficiency
Model Size	80-180 MB (recognition)	10-16 MB (recognition)
Inference Speed (GPU)	8-9 ms	5-6 ms
Inference Speed (CPU)	30-40 ms	17-21 ms
Accuracy	Higher (86%+)	Good (81%+)
Target Deployment	Servers, cloud	Edge devices, mobile
Memory Requirements	Higher	Lower

Detection Model Comparison:

Model	Hmean (%)	Model Size (MB)	GPU Time (ms)	CPU Time (ms)
PP-OCRv5_server_det	83.8	84.3	89.55 / 70.19	383.15 / 383.15
PP-OCRv5_mobile_det	79.0	4.7	10.67 / 6.36	57.77 / 28.15

Recognition Model Comparison:

Model	Avg Acc (%)	Model Size (MB)	GPU Time (ms)	CPU Time (ms)
PP-OCRv5_server_rec	86.38	81	8.46 / 2.36	31.21 / 31.21
PP-OCRv5_mobile_rec	81.29	16	5.43 / 1.46	21.20 / 5.32

Sources: docs/version3.x/pipeline_usage/OCR.md118-169 docs/version3.x/pipeline_usage/OCR.en.md118-168

3.2 Scale Selection Strategy

Configuration Examples:

Server deployment (maximum accuracy):

Mobile deployment (efficiency optimized):

Sources: docs/version3.x/pipeline_usage/OCR.md132-200 docs/quick_start.en.md68-80

4. Language Support

4.1 Language Coverage and Model Registry Architecture

PaddleOCR provides three-tiered language support with distinct model naming conventions:

Title: Language Support Architecture and Model Registry Names

Model Naming Convention:

• Pattern: {lang}_PP-OCRv{version}_{scale}_rec
• Examples:
  • PP-OCRv5_server_rec (default multilingual)
  • korean_PP-OCRv5_mobile_rec (Korean-specific)
  • en_PP-OCRv4_mobile_rec (English-specific PP-OCRv4)

Language Selection Strategy:

Language(s)	Recommended Model	Registry Name	Size
Chinese + English	PP-OCRv5 default	PP-OCRv5_server_rec or PP-OCRv5_mobile_rec	81MB / 16MB
Korean	Korean-specific	korean_PP-OCRv5_mobile_rec	14MB
Latin languages (French, Spanish, Portuguese, etc.)	Latin model	latin_PP-OCRv5_mobile_rec	14MB
Arabic script	Arabic model	arabic_PP-OCRv5_mobile_rec	7.6MB
Thai	Thai model	th_PP-OCRv5_mobile_rec	7.5MB
Multiple diverse languages	Vision-language model	PaddleOCR-VL pipeline	900MB

Sources: docs/version3.x/pipeline_usage/OCR.md242-632 docs/version3.x/pipeline_usage/OCR.en.md241-631 docs/version3.x/pipeline_usage/OCR.md432-531

4.2 PP-OCRv5 Multi-Type Single Model

The default PP-OCRv5_rec models support five text types within a single model, eliminating the need to switch models for mixed-language documents:

Model	Chinese	English	Traditional Chinese	Japanese	Model Size
PP-OCRv5_server_rec	86.38%	64.70%	93.29%	60.35%	81 MB
PP-OCRv5_mobile_rec	81.29%	66.00%	83.55%	54.65%	16 MB

Use Cases:

• Documents with mixed Chinese-English text
• Japanese documents containing Chinese characters
• Pinyin annotations in educational materials
• Traditional Chinese documents with modern annotations

Configuration:

Sources: docs/version3.x/pipeline_usage/OCR.md246-285 docs/version3.x/pipeline_usage/OCR.en.md245-284

4.3 Language-Specific Recognition Models

For languages beyond the five types in PP-OCRv5, PaddleOCR provides specialized recognition models optimized for specific language families:

4.3.1 Available Language Models

PP-OCRv5 Series (Latest):

Language/Script	Model Name	Accuracy	Languages Covered
Korean	korean_PP-OCRv5_mobile_rec	88.0%	Korean, English, Numbers
Latin	latin_PP-OCRv5_mobile_rec	84.7%	Most Latin-based languages
Eastern Slavic	eslav_PP-OCRv5_mobile_rec	81.6%	Russian, Ukrainian, Belarusian
Thai	th_PP-OCRv5_mobile_rec	82.68%	Thai, English, Numbers
Greek	el_PP-OCRv5_mobile_rec	89.28%	Greek, English, Numbers
Arabic	arabic_PP-OCRv5_mobile_rec	81.27%	Arabic script languages
Cyrillic	cyrillic_PP-OCRv5_mobile_rec	80.27%	All Cyrillic-based languages
Devanagari	devanagari_PP-OCRv5_mobile_rec	84.96%	Hindi, Sanskrit, etc.
Telugu	te_PP-OCRv5_mobile_rec	87.65%	Telugu, Numbers
Tamil	ta_PP-OCRv5_mobile_rec	94.2%	Tamil, Numbers
English	en_PP-OCRv5_mobile_rec	85.25%	English (improved accuracy)

PP-OCRv3 Series (Legacy):

Additional languages available through PP-OCRv3 models include Japanese (japan_PP-OCRv3_mobile_rec), Kannada (ka_PP-OCRv3_mobile_rec), and others with model sizes around 8-10 MB.

Sources: docs/version3.x/pipeline_usage/OCR.md422-632 docs/version3.x/pipeline_usage/OCR.en.md422-631

4.3.2 Language Model Selection

Configuration Example:

Sources: docs/version3.x/pipeline_usage/OCR.md432-531 docs/version3.x/pipeline_usage/OCR.en.md388-531

4.4 PaddleOCR-VL: Universal Language Support

For comprehensive multilingual support beyond PP-OCRv5's capabilities, PaddleOCR-VL provides a unified 0.9B parameter vision-language model supporting 111 languages:

Key Characteristics:

• Single model for 111 languages including rare languages (Tibetan, Bengali)
• Unified architecture: NaViT visual encoder + ERNIE-4.5-0.3B LLM
• Supports complex elements: text, tables, formulas, charts
• Document parsing optimized for real-world scenarios

When to Use PaddleOCR-VL:

• Documents with rare or unsupported languages
• Complex document parsing requirements
• Need for integrated layout and content understanding
• Willingness to use larger model (900M parameters vs. 16-81M for PP-OCRv5)

Configuration:

Sources: README.md61-66 README.md89-98 docs/index.en.md27-29

5. Model Configuration and Selection API

5.1 Configuration Parameters and Model Registry

The PaddleOCR class paddleocr/__init__.py exposes model selection through these parameters:

Title: PaddleOCR Configuration Parameters for Model Selection

Configuration Parameter Details:

Parameter	Type	Valid Values	Default	Resolution Logic
ocr_version	str	"PP-OCRv5", "PP-OCRv4", "PP-OCRv3"	"PP-OCRv5"	Sets version prefix for model names
text_detection_model_name	str	Model registry name (e.g., "PP-OCRv5_server_det")	"PP-OCRv5_server_det"	Direct model name lookup
text_recognition_model_name	str	Model registry name (e.g., "korean_PP-OCRv5_mobile_rec")	"PP-OCRv5_server_rec"	Direct model name lookup
lang	str	Language code (e.g., "korean", "arabic")	"ch"	Maps to language-specific model if available

Automatic Model Selection Logic:

When lang is specified without explicit text_recognition_model_name, the system performs automatic model resolution:

Manual Model Selection (highest priority):

Model Download and Caching:

Models are downloaded on first use to:

• Location: ~/.paddleocr/ or PADDLE_PDX_MODEL_SOURCE environment variable path
• Format: .pdmodel (model structure) + .pdiparams (weights) + config files
• Automatic: Downloaded from https://paddle-model-ecology.bj.bcebos.com/paddlex/ on first invocation

Sources: docs/version3.x/pipeline_usage/OCR.md746-1048 docs/version3.x/pipeline_usage/OCR.en.md766-1048 pyproject.toml41-45

5.2 Command Line Interface (CLI) Model Selection

The paddleocr command-line interface paddleocr/__main__.py exposes model selection through these flags:

Command: paddleocr ocr [OPTIONS]

Model Selection Flags:

CLI Flag	Type	Example Value	Equivalent Python Parameter
--ocr_version	str	PP-OCRv5, PP-OCRv4, PP-OCRv3	ocr_version
--text_detection_model_name	str	PP-OCRv5_mobile_det	text_detection_model_name
--text_recognition_model_name	str	korean_PP-OCRv5_mobile_rec	text_recognition_model_name
--lang	str	korean, arabic, ch	lang

CLI Usage Examples:

Output Format (saved to --save_path if specified):

• Text results in .txt files
• Detection boxes visualization
• Full pipeline results in JSON format

Sources: docs/version3.x/pipeline_usage/OCR.md761-810 docs/version3.x/pipeline_usage/OCR.en.md766-825 paddleocr/__main__.py

5.3 Model Selection Decision Matrix

Decision Guidelines:

1. Start with PP-OCRv5 defaults: PP-OCRv5_server_det + PP-OCRv5_server_rec
2. Switch to mobile: If deploying on edge devices or need < 20ms latency
3. Use language-specific: If primary language is not Chinese/English/Japanese
4. Use PaddleOCR-VL: If need 100+ languages or complex document parsing
5. Use PP-OCRv4: Only for legacy compatibility or specific character set needs (15K+)

Sources: docs/version3.x/pipeline_usage/OCR.md701 README.md67-75

6. Model Performance Comparison

6.1 Cross-Version Performance

Detection models across versions (server variants):

Version	Model	Hmean (%)	GPU Time (ms)	Size (MB)
v5	PP-OCRv5_server_det	83.8	89.55	84.3
v4	PP-OCRv4_server_det	69.2	127.82	109
v3	N/A (mobile only)	-	-	-

Recognition models across versions (mobile variants):

Version	Model	Avg Acc (%)	GPU Time (ms)	Size (MB)
v5	PP-OCRv5_mobile_rec	81.29	5.43	16
v4	PP-OCRv4_mobile_rec	78.74	5.26	10.5
v3	PP-OCRv3_mobile_rec	72.96	3.89	10.3

Key Improvements in PP-OCRv5:

• 14.6% detection accuracy improvement (v4 → v5)
• 13% recognition accuracy improvement across multiple scenarios
• Single model for 5 text types (previously required separate models)
• Better handling of handwriting, vertical text, and rare characters

Sources: docs/version3.x/pipeline_usage/OCR.md132-169 docs/version3.x/pipeline_usage/OCR.md287-334

6.2 Language Model Performance

Performance of language-specific PP-OCRv5 models (all mobile variants, 5.43ms GPU inference):

Language	Model	Accuracy (%)	Model Size (MB)
Tamil	ta_PP-OCRv5_mobile_rec	94.2	7.5
Greek	el_PP-OCRv5_mobile_rec	89.28	7.5
Korean	korean_PP-OCRv5_mobile_rec	88.0	14
Telugu	te_PP-OCRv5_mobile_rec	87.65	7.5
English	en_PP-OCRv5_mobile_rec	85.25	7.5
Devanagari	devanagari_PP-OCRv5_mobile_rec	84.96	7.5
Latin	latin_PP-OCRv5_mobile_rec	84.7	14
Thai	th_PP-OCRv5_mobile_rec	82.68	7.5
Eastern Slavic	eslav_PP-OCRv5_mobile_rec	81.6	14
Arabic	arabic_PP-OCRv5_mobile_rec	81.27	7.6
Cyrillic	cyrillic_PP-OCRv5_mobile_rec	80.27	7.7

All models maintain efficient inference speeds (5.43ms GPU, 21.20ms CPU) with compact sizes (7-14 MB).

Sources: docs/version3.x/pipeline_usage/OCR.md432-531

7. Best Practices and Model Selection Workflows

7.1 Model Selection Decision Tree

Title: Model Selection Decision Tree with Registry Names

Model Selection Checklist:

1. Language Coverage → Select model tier (PP-OCRv5 default, language-specific, or PaddleOCR-VL)
2. Deployment Environment → Choose scale (_server_ vs. _mobile_)
3. Resource Budget → Validate model size fits memory constraints
4. Accuracy Requirements → Enable high-performance inference if needed
5. Validation → Benchmark on representative dataset before production

Sources: docs/version3.x/pipeline_usage/OCR.md701-1048 docs/version3.x/pipeline_usage/OCR.en.md766-1048

7.2 Common Configuration Patterns with Code Examples

Pattern 1: Default Server Deployment (Chinese-English Mixed Documents)

Pattern 2: Mobile Deployment (Korean Text, Resource-Constrained)

Pattern 3: High-Accuracy Server with TensorRT Acceleration

Pattern 4: Document-Optimized Recognition (15K+ Characters)

Pattern 5: Multi-Language Global Support (111 Languages)

Pattern 6: Latin Languages (French, Spanish, Portuguese, etc.)

Pattern 7: CLI Batch Processing with Language Auto-Selection

Sources: docs/version3.x/pipeline_usage/OCR.md746-1048 docs/version3.x/pipeline_usage/OCR.en.md766-1048 docs/quick_start.en.md66-94

7.3 Performance Optimization Guidelines

1. For Maximum Accuracy:

   • Use PP-OCRv5_server_det + PP-OCRv5_server_rec
   • Enable high-performance inference (enable_hpi=True)
   • Use TensorRT acceleration on supported GPUs
   • Consider preprocessing modules for challenging images

2. For Maximum Speed:

   • Use PP-OCRv5_mobile_det + PP-OCRv5_mobile_rec
   • Disable optional preprocessing (use_doc_orientation_classify=False, use_doc_unwarping=False)
   • Reduce limit_side_len parameter for faster processing
   • Deploy with MKL-DNN on CPU or TensorRT on GPU

3. For Minimal Memory Footprint:

   • Use mobile variants (16-81 MB vs. 81-173 MB for server)
   • Language-specific models when single language needed (7-14 MB)
   • Disable unused optional modules

4. For Mixed-Language Documents:

   • First try PP-OCRv5_server_rec (supports 5 text types)
   • If languages not covered, use PaddleOCR-VL for comprehensive support
   • Avoid switching between language-specific models (slower)

Sources: docs/version3.x/pipeline_usage/OCR.md707 README.md191-199

7.4 Migration from Legacy Versions (PaddleOCR 2.x to 3.x)

API Parameter Changes:

PaddleOCR 2.x Parameter	PaddleOCR 3.x Parameter	Notes
det_model_dir	text_detection_model_name	Now uses model registry names instead of paths
rec_model_dir	text_recognition_model_name	Automatic download from model hub
cls_model_dir	textline_orientation_model_name	For text line orientation
use_angle_cls	use_textline_orientation	Renamed for clarity
lang	lang	Unchanged, but now supports 20+ languages

Migration Examples:

Key Migration Benefits:

1. Automatic Model Management: Models downloaded from central registry (no manual download)
2. Unified Multilingual Support: Single PP-OCRv5_server_rec handles Chinese, English, Traditional Chinese, Japanese
3. Improved Accuracy: 13% average accuracy improvement with PP-OCRv5
4. Simplified Configuration: Model names replace paths

PaddleX Dependency (Transparent to Users):

PaddleOCR 3.x uses PaddleX for underlying inference pyproject.toml42-45:

• Dependency: paddlex[ocr-core]>=3.4.0,<3.5.0
• Optional Features: paddlex[doc-parser], paddlex[ie], paddlex[trans]
• User Impact: Transparent; no direct PaddleX knowledge required for basic usage

Backward Compatibility Notes:

• Old parameter names (det_model_dir, etc.) deprecated but may still work with warnings
• Model files from 2.x (*.pdmodel, *.pdiparams) can still be loaded via text_detection_model_dir parameter (not text_detection_model_name)
• Recommended to migrate to 3.x registry-based model names for maintainability

Sources: docs/version3.x/paddleocr_and_paddlex.md1-50 docs/version3.x/paddleocr_and_paddlex.en.md1-50 pyproject.toml41-46

---

Sources for entire document: README.md1-246 docs/version3.x/pipeline_usage/OCR.md1-810 docs/version3.x/pipeline_usage/OCR.en.md1-810 docs/version3.x/pipeline_usage/PP-StructureV3.md1-320 docs/version3.x/pipeline_usage/PP-StructureV3.en.md1-320 docs/index.en.md1-92 docs/quick_start.en.md1-140