Summary

Generates training data with JSON annotations for training Tesseract OCR on custom text characters/codes and is fully compatible with Supervisely.

COMING SOON:

1. GUI app to generate/edit box and tiff files for Tesseract 4.0.0 training
2. One-click training as well at some point.

Output

Generates the outputdir directory with the following tree structure:

├── outputdir
│   ├── train
│   │   ├── img
│   │   ├── ann
│   ├── test
│   │   ├── img
│   │   ├── ann
│   ├── meta.json

With images stored in their respective train (or test) /img folder and annotations in the train (or test) /ann folder. The meta.json file contains data for Supervisely. The annotations Json contains the text in the image, and the Top Left and Bottom Right co-ordinates of the bounding box. The text can be accessed by json['objects'][0]['description'] and the points can be accessed by json['objects'][0]['points']['exterior'].

How to use

1. Clone the repo with git clone https://github.com/rafayk7/tesseractDataGenerator.git
2. Download requirements with pip install requirements.txt
3. Run run.py with python3 run.py

Parameters to change

In run.py

1. trainingAmt - Number of total images to be generated (default 1000)
2. trainTestSplit - Split between number of training and testing images in range [0,1] (default 0.7)
3. codesfile - path to the file with codes/chars to generate training images from (default codesKOR.txt)
4. outputdir - directory to store all generated data (default /data)

In Utils.py

1. getTextImgForTraining
   1. minAngle - lower bound angle for skew generation
   2. maxAngle - upper bound angle for skew generation
   3. (W, H) - size of text image (not final training image - see getTrainingImage)
   4. fontSize - size of font of text - change with getTrainingImage
2. getImgJson
   1. format of JSON annotations file - currently in compatibility with Supervisely
   2. width, height - width and height of training image, change with (Wimg, Himg) in getTrainingImage
3. getTrainingImg
   1. fontSize - size of font of text - change with getTextImgForTraining
   2. (Wimg, Himg) - size of final training image
   3. backgroundImgPath - path to background image for training images

Sample Results