neots 0.1.2

NeoTextSynthesizer - High-performance OCR training data generator

NeoTextSynthesizer (neots)

A high-performance, multi-threaded synthetic OCR training data generator built with C++.

NeoTextSynthesizer generates realistic text images with diverse fonts, backgrounds, and visual effects — ideal for training Text Recognition models.

Features

• High Performance: Multi-threaded C++ core with shared glyph cache, background image cache, and lock-free rendering pipeline
• Multi-language Support: Built-in support for Chinese (Simplified & Traditional), Japanese, Korean, English, German, French, Italian, Spanish, Russian, and extensible to any language
• Flexible Text Sampling: Sequential corpus streaming and random character sampling with nested probability-weighted groups
• Rich Visual Effects: Random fonts, font sizes, scale variations, italic/underline/strikethrough effects, diverse backgrounds (solid colors, gradients, real images)
• Python Integration: Full Python API via nanobind with NumPy/PIL support
• CLI Tools: neots for workspace initialization and batch generation, flat-wiki for Wikipedia corpus generation
• Cross-platform: Supports Linux x86_64, Linux AArch64, macOS ARM64, and Windows x86_64

Installation

Install pre-built wheels from PyPI:

pip install neots

Quick Start

CLI Usage

# Initialize a workspace
neots init

# Generate 10,000 images
neots generate --total 10000

# Generate with custom config and workers
neots generate --config config.yaml --total 50000 --workers 8

# Show help
neots help

Wikipedia Corpus Generation (CLI)

# Generate a Chinese corpus from Wikipedia
flat-wiki --lang zh --output ./corpuses/zh.txt --min_len 5 --sample_count 10000

# Generate with vocabulary filter
flat-wiki -l en -o ./corpuses/en.txt -v dictionary.txt -n 5000

Python API

from neots import NeoTextSynthesizer

# Initialize from config file
synth = NeoTextSynthesizer.from_config_file("config.yaml")

# Initialize from a dictionary
config = NeoTextSynthesizer.get_default_config()
config["generate"]["output_height"] = 64
synth = NeoTextSynthesizer.from_config(config)

# Initialize from built-in default preset
synth = NeoTextSynthesizer.from_default()          # full preset
synth = NeoTextSynthesizer.from_default("minimal")  # minimal preset

# Batch generate (blocks until complete)
synth.generate(total=10000, workers=8, show_progress=True)

# Generate single image as PIL Image
img = synth.generate_instance("Hello World", type="PIL")
img.save("sample.png")

# Generate as NumPy array (H, W, 3) RGB
arr = synth.generate_instance("测试文本", type="numpy")

# Save directly to file
synth.generate_instance("サンプル", type="file", path="output.png")

# Export default config
NeoTextSynthesizer.output_default_dict("my_config.yaml", type="yaml")
NeoTextSynthesizer.output_default_dict("my_config.json", type="json")

Corpus Generation (Python API)

from neots import FlattenWikipediaGenerator

gen = FlattenWikipediaGenerator(
    lang="zh",                        # Language code
    output="./corpuses/zh.txt",       # Output file path
    vocab="dictionary_extended.txt",  # Optional: filter by vocabulary
    min_len=5,                        # Minimum line length
    sample_count=10000,               # Number of articles to process
)
gen.generate()

Configuration

Configuration is structured into three sections. See config.yaml for a complete example after initialization.

text_sampler

Defines how text strings are sampled for rendering. Each entry has a probability weight and a data source.

Type	Description
sequential	Streams text from a flat file. Ideal for natural language corpuses.
random	Randomly samples from character files or nested sub-groups.

text_sampler:
  # Stream Chinese text from a corpus file
  - type: sequential
    prob: 0.20
    from_file: ./corpuses/zh.txt

  # Random character sampling with nested groups
  - type: random
    prob: 0.20
    sub_items:
      - prob: 0.4
        from_file: ./symbols_characters/characters.txt
        section: [~, 3500]    # First 3500 lines only
      - prob: 0.1
        from_string: " "      # Inline source

Key parameters:

• prob: Probability weight (auto-normalized)
• from_file / from_string: Data source
• section: [start, end]: Slice of lines (~ = null = beginning/end)
• sub_items: Nested probability groups (recursive)
• traditional_prob: Probability of simplified → traditional Chinese conversion

image_processor

Controls visual appearance:

Parameter	Default	Description
bg_image_prob	0.3	Probability of using a real background image
gray_bg_prob	0.7	Probability of gray vs. colored background
effect_prob	0.2	Probability of text effects (italic/underline/strikethrough)
font_size	55	Base font size in pixels
scale_range	[0.7, 1.1]	Random scale factor range
margin_range	[-5, 20]	Horizontal margin range (pixels)
offset_prob	0.5	Probability of random position offset

generate

Controls output structure:

Parameter	Default	Description
min_targets / max_targets	5 / 50	Character count range per image
output_height	48	Output image height (width auto-calculated)
default_font_dir	./default_fonts	Primary fonts directory
fallback_font_dir	./fallback_fonts	Fallback fonts directory
out_dir	./generated	Output image directory
out_jsonl	./generated.jsonl	Annotation file path
hierarchical_structure	[100, 625]	Directory nesting levels

Font Setup

Primary Fonts (default_fonts/)

Place diverse .ttf / .otf font files here. A random font is selected for each image. More fonts = better model generalization. Aim for 50-200+ fonts covering your target scripts.

Fallback Fonts (fallback_fonts/)

When the primary font doesn't cover certain characters, the system falls back to these fonts. Place comprehensive Unicode fonts here (e.g., Noto Sans CJK, Noto Sans).

System Default

If no fallback fonts are provided, the system automatically tries:

• Linux: /usr/share/fonts/truetype/DejaVuSans.ttf
• macOS: /System/Library/Fonts/Courier New.ttf
• Windows: C:\Windows\Fonts\Arial.ttf

If the system font is also unavailable, initialization fails.

Background Images

For realistic training data, add images to backgrounds/ and set bg_image_prob > 0.

Supported formats: .jpg, .jpeg, .png, .bmp, .webp

Output Format

Generated images are saved as PNG files in a hierarchical directory structure:

generated/
├── 00000001/
│   ├── 00000001/
│   │   ├── 00000001.png
│   │   ├── 00000002.png
│   │   └── ...
│   └── ...
└── ...

Annotations are saved in JSONL format (generated.jsonl):

{"image":"00000001/00000001/00000001.png","text":"示例文本","length":248}
{"image":"00000001/00000001/00000002.png","text":"Sample text","length":315}

Each line contains:

• image: Relative path to the image file
• text: The rendered text content
• length: Image width in pixels

Development

This project uses vcpkg to manage native C++ dependencies, and uses scikit-build-core + nanobind for the Python/C++ bridge. Development is now based on configuring the environment with vcpkg and building from source.

Supported Platforms

• Linux x86_64
• Linux AArch64
• macOS ARM64
• Windows x86_64

Build from Source with vcpkg

Windows

git clone https://github.com/microsoft/vcpkg.git
.\vcpkg\bootstrap-vcpkg.bat
$env:CMAKE_TOOLCHAIN_FILE="$PWD\vcpkg\scripts\buildsystems\vcpkg.cmake"
pip install -e . --no-build-isolation

macOS

git clone https://github.com/microsoft/vcpkg.git
./vcpkg/bootstrap-vcpkg.sh
export CMAKE_TOOLCHAIN_FILE="$PWD/vcpkg/scripts/buildsystems/vcpkg.cmake"
pip install -e . --no-build-isolation

Linux

git clone https://github.com/microsoft/vcpkg.git
./vcpkg/bootstrap-vcpkg.sh
export CMAKE_TOOLCHAIN_FILE="$PWD/vcpkg/scripts/buildsystems/vcpkg.cmake"
pip install -e . --no-build-isolation

If you prefer to build the native extension directly with CMake, make sure CMAKE_TOOLCHAIN_FILE points to the vcpkg toolchain before configuring the project.

Donate

This project is maintained by a single person, so the documentation is not yet fully comprehensive and the feature set is still relatively limited.

If you would like to support the project, the most valuable help is:

• submitting pull requests
• actively opening issues when you encounter problems during usage

License

This project is licensed under the MIT License. See LICENSE for details.