The keras-ocr package generally consists of two parts: a Detector and a Recognizer:

• Detector is responsible for creating bounding boxes for the words of the text.
• Recognizer is responsible for processing batch of cropped parts of the initial image.

Keras-ocr connects this two parts into seamless pipeline. "Out of the box", it can handle a wide range of images with texts. But in a specific task, when the field of possible images with texts is greatly narrowed, it shows itself badly in the Recognizer part of the task.

In this regard, the task of fine-tuning Recognizer on a custom dataset was set.

$ python3 -m venv keras_ocr
$ pip install keras-ocr

And TRDG library for synthetic text generation.

$ pip install trdg

We will use the TRDG library to generate synthetic text. All necessary code presented in the data_generation.py. Things you need to know:

• You choose template for generating text, e.g. if template is "({}{}/{})", then all brackets will be randomly filled with symbols from alphabet. You need to specify your own instance of StringTemplate classs.

• You choose the alphabet. In our example case it contains only digits. P.S. Some of the repeated in data_generation.py, hence emperical distribution probability for each symbol defined as fraction of n_repeats to alphabet_size.

• You can choose your own fonts. To do this, follow instruction:

  1. Download needed fonts as .ttf files
  2. Go to trdg fonts directory ./keras_ocr/lib/python3.8/site-packages/trdg/fonts/
  3. Create directory $ mkdir cs (cs means custom fonts), you can chooce the disered name
  4. Place fonts files in this dir
  5. (For Mac users only) Don't forget to remove .DS_Store from this folder

• You can chooce image background for text. When creating instance of GeneratorFromStrings in function generate_data_units(...), provide folder with images with arg image_dir

data_generator = DataGenerator(string_templates=[StringTemplate('{}{}{}{}{}{}{}', 7)])

data_generator.generate(n_patches=20000, n_total_samples=550, path='DigitsBracketsDataset/train')

• n_patches -- number of different strings from provided template
• n_total_samples -- number of total samples from patches
• path -- dir to save samples

Follow instruction in fine_tuning.ipynb. Don't forget to add function get_custom_dataset(...) to datasets.py in keras-ocr package directory (./keras_ocr/lib/python3.8/site-packages/keras_ocr/datasets.py):

def get_custom_dataset(path: str, split: str):
    """
    param: path: path to dataset root dir (include train/test dirs)
    Returns:
        A recognition dataset as a list of (filepath, box, word) tuples
    """
    data = []
    if split == 'train':
        train_dir = os.path.join(path, 'train')
        data.extend(
            _read_born_digital_labels_file(
                labels_filepath=os.path.join(train_dir, "gt.txt"),
                image_folder=train_dir,
            )
        )
    elif split == 'test':
        test_dir = os.path.join(path, 'test')
        data.extend(
            _read_born_digital_labels_file(
                labels_filepath=os.path.join(test_dir, 'gt.txt'), 
                image_folder=test_dir
            )
        )
    return data