OCR4All Pixel Classifier

Requirements

Python dependencies are specified in requirements.txt / setup.py.

The package is tested with Tensorflow 2.0 up to 2.5. If you want to use a GPU, you'll have to set up your system with the CUDA and CuDNN versions matching your used Tensorflow version. If using Tensorflow older than 2.1 for some reason, you'll additionaly have to replace the tensorflow package with tensorflow-gpu manually.

Usage

For training and direct usage, install ocr4all-pixel-classifier-frontend. This package only contains the library code.

Pixel classifier

Classification

To run a model on some input images, use ocr4all-pixel-classifier predict:

ocr4all-pixel-classifier predict --load PATH_TO_MODEL \
	--output OUTPUT_PATH \
	--binary PATH_TO_BINARY_IMAGES \
	--images PATH_TO_SOURCE_IMAGES \
	--norm PATH_TO_NORMALIZATIONS

(ocr4all-pixel-classifier is an alias for ocr4all-pixel-classifier predict)

This will create three folders at the output path:

• color: the classification as color image, with pixel color corresponding to the class for that pixel
• inverted: inverted binary image with classification of foreground pixels only (i.e. background is black, foreground is white or class color)
• overlay: classification image layered transparently over the original image

Training

For training, you first have to create dataset files. A dataset file is a JSON file containing three arrays, for train, test and evaluation data (also called train/validation/test in other publications). The JSON file uses the following format:

{
	"train": [
		//datasets here
	],
	"test": [
		//datasets here
	],
	"eval": [
		//datasets here
	]
}

A dataset describes a single input image and consists of several paths: the original image, a binarized version and the mask (pixel color corresponds to class). Furthermore, the line height of the page in pixels must be specified:

{
	"binary_path": "/path/to/image/binary/filename.bin.png",
	"image_path":  "/path/to/image/color/filename.jpg",
	"mask_path":  "/path/to/image/mask/filename_MASK.png",
	"line_height_px": 18
}

The generation of dataset files can be automated using ocr4all-pixel-classifier create-dataset-file. Refer to the command's --help output for further information.

To start the training:

ocr4all-pixel-classifier train \
    --train DATASET_FILE.json --test DATASET_FILE.json --eval DATASET_FILE.json \
    --output MODEL_TARGET_PATH \
    --n_iter 5000

The parameters --train, --test and --eval may be followed by any number of dataset files or patterns (shell globbing).

Refer to ocr4all-pixel-classifier train --help for further parameters provided to affect the training procedure.

You can combine several dataset files into a split file. The format of the split file is:

{
	"label": "name of split",
	"train": [
		"/path/to/dataset1.json",
		"/path/to/dataset2.json",
		...
	],
	"test": [
		//dataset paths here
	],
	"eval": [
		//dataset paths here
	]
}

To use a split file, add the --split_file parameter.

Examples

See the examples for dataset generation and training

ocr4all-pixel-classifier compute-image-normalizations / ocrd_compute_normalizations

Calculate image normalizations, i.e. scaling factors based on average line height.

Required arguments:

• --input_dir: location of images
• --output_dir: target location of norm files

Optional arguments:

• --average_all: Average height over all images
• --inverse