belval/textrecognitiondatagenerator

This tool generates fake images of text that can be used to train AI systems to read printed or handwritten words, a task called OCR (optical character recognition). Training a good OCR system requires thousands of example images paired with the correct text label, and collecting real examples is slow and expensive. This tool creates them synthetically, letting you generate as many as you need in seconds. You give the tool a language and some parameters, and it picks words from a built-in dictionary, renders them using fonts from a folder of your choice, and saves the resulting images along with their text labels. You can control a wide range of visual properties: font size, whether the text is skewed or distorted, how much blurring is applied, what the background looks like (noise, white, a patterned texture, or a custom photo), the spacing between characters, and even stroke width. There is also an experimental mode that produces images resembling handwritten text. The tool supports Latin-alphabet languages by default and also works with Chinese (both simplified and traditional), Japanese, Korean, Thai, and others. Adding support for a new language requires dropping a font file and a word-list dictionary into the right folders, which takes only a few minutes. You can run it from the command line or import it as a Python library to slot it directly into your model training pipeline. A Docker image is also available if you prefer not to install any dependencies locally. The benchmark figures in the README show it can produce over 3,000 images per second on a modern multi-core machine when using multiple threads.