OCR Text Recognition

iClick provides built-in OCR (Optical Character Recognition) functionality that can meet the text recognition needs of most automation scenarios. This article explains the advantages of built-in OCR and why we don't integrate third-party OCR libraries like PaddleOCR.

How It Works

iClick's built-in OCR performs recognition on the iPhone device, not on the computer host.

Core Advantages:

• ✅ Multi-language Support: Based on iOS system capabilities, supports 18+ major languages worldwide without additional configuration
• ✅ Zero Host Resource Usage: Recognition process completed on iOS device, does not consume computer CPU, memory, or GPU resources

Applicable Scenarios

Built-in OCR can meet most automation needs:

• ✅ App interface text recognition / button and menu text extraction
• ✅ Message notification content recognition / in-game text recognition
• ✅ Simple data collection tasks

Recommendation

If your needs do not involve large-scale document scanning, built-in OCR is completely sufficient, no need to consider third-party OCR libraries.

Why Not Integrate Third-party OCR (Using Baidu PaddleOCR as Example)

1. Model Files Too Large, Integrating Single Language Makes No Sense

PaddleOCR and other deep learning OCR library model files are very large:

• Single language model: typically 100MB - 500MB
• Multi-language support: requires downloading multiple models, easily exceeding 1GB - 3GB
• High-precision models: can reach 5GB or larger
• Each language service requires separate model files

2. GPU Version Requirements Are Demanding

Hardware Limitations

• ⚠️ Only supports NVIDIA graphics cards (AMD cards, integrated graphics cannot be used)
• ⚠️ Requires mid to high-end cards: RTX 2060 or higher
• ⚠️ Official CPU version exists but efficiency makes it impractical

Performance Bottleneck

Even with a dedicated graphics card, performance is very limited. Example with RTX 3060 + 15 iPhones for button recognition:

Configuration: NVIDIA GeForce RTX 3060 (12GB)
Devices: 15 iPhones
Optimization: Extreme concurrency control and performance optimization
Result:
  - GPU already running at full capacity (100% usage)
  - OCR recognition speed unbearably slow
  - Overall performance severely degraded

You Can Still Use Third-party OCR

If you still need to use third-party OCR libraries like PaddleOCR (such as for doc scan, specific language recognition, etc.), and you have a high-performance NVIDIA graphics card, you can integrate it yourself.

Advantages of Self-integration

In fact, building PaddleOCR and other OCR libraries into independent services is very simple:

• 🚀 Minimal Code - Building a basic OCR HTTP service requires only dozens of lines of code
• 🔧 Flexible Control - Can choose desired language models, customize features and parameters
• 🎯 Specialized Optimization - Optimize models for specific scenarios (such as captchas)

Applicable Scenarios

Third-party OCR is suitable for the following special scenarios:

Scenario	Recommended Solution	Description
Captcha Recognition	Specialized captcha model	Targeted training, higher recognition rate
Specific Languages	Corresponding language model	Such as minority languages, dialects
Handwritten Text	Handwriting recognition model	Built-in OCR has limited handwriting support
Table Recognition	Table-specific model	Structured data extraction
Document Scanning	High-precision model	Bulk document processing