Belval/TextRecognitionDataGenerator
TextRecognitionDataGenerator
TextRecognitionDataGenerator is a system for creating synthetic text images and character masks to train and test optical character recognition models. It functions as a dataset tool that generates artificial images by combining dictionary words with specific fonts and backgrounds to produce training sets for machine learning.
The project includes a handwriting simulation engine that uses machine learning models to produce synthetic handwritten text. It also features a document degradation simulator that applies skew, blur, and distortion to images to mimic the visual quality of scanned physical documents.
The tool covers a wide range of capabilities, including multilingual text recognition for scripts such as Latin, Chinese, Japanese, Korean, and Thai. It supports text appearance customization and multi-source text integration from dictionaries or external APIs, while providing character-level masks to establish ground truth for model training.
Features
- Synthetic Training Data Generators - Produces artificial text images and ground-truth annotations specifically to train and test optical character recognition models.
- OCR Dataset Generators - Generates artificial text images and corresponding character-level masks to create comprehensive training datasets for OCR models.
- OCR Dataset Toolkits - Provides a comprehensive system for generating diverse multilingual training sets using customizable fonts and dictionary-based text sampling.
- Binary Mask Generators - Produces binary masks at the pixel level to isolate characters from backgrounds for machine learning training.
- Handwriting Image Generators - Generates synthetic images of human-like handwriting using a trained machine learning model.
- Handwriting Synthesis Models - Uses pre-trained neural networks to synthesize realistic handwritten image sequences from digital text.
- OCR Model Training - Provides large synthetic image datasets with labels to train neural networks for visual text recognition.
- Multilingual Text Recognition - Builds training datasets that support a wide variety of global scripts, including Latin, CJK, and Thai.
- Multilingual Font Loading - Supports the loading of specific font directories to enable text recognition for Latin, Chinese, Japanese, Korean, and Thai scripts.
- Synthetic Image Generators - Programmatically creates images of text by combining dictionary words with specific fonts and backgrounds.
- OpenType Font Rendering - Implements high-quality glyph mapping and rendering using OpenType and Unicode standards to support diverse scripts.
- Character Ground-Truth Masks - Generates precise binary image masks for each character to establish ground truth for training OCR models.
- Handwriting Recognition Training - Creates synthetic handwritten samples to improve the training and accuracy of handwriting recognition models.
- Distortion Simulation - Applies skewing and warping operations to simulate the geometric perspective and distortions of scanned physical documents.
- Blur Simulations - Simulates low-resolution optical capture and defocused effects by applying Gaussian blur to generated text images.
- Scanned Document Simulation - Generates distorted or blurred text images that mimic the visual quality and degradation of scanned physical documents.
- Degradation Simulation Pipelines - Integrates blur, skew, and distortion into a pipeline to simulate the quality degradation of scanned documents.
- Typographic Visual Styling - Provides control over visual properties including stroke width, color, character spacing, and layout margins.
