page-segmentation module for OCR-d

Introduction

This module implements a page segmentation algorithm based on a Fully Convolutional Network (FCN). The FCN creates a classification for each pixel in a binary image. This result is then segmented per class using XY cuts.

Requirements

For GPU-Support: CUDA and CUDNN
other requirements are installed via Makefile / pip, see requirements.txt in repository root.

Installation

make deps

to install dependencies and

make install

to install the package.

Both install python packages via pip, so you may want to activate a virtalenv before installing.

If you share a virtualenv with a package requiring tensorflow < 2.1 and want to use a GPU, replace the tensorflow package with tensorflow-gpu manually.

Usage

ocrd-pc-segmentation follows the ocrd CLI.

It expects a binary page image and produces region entries in the PageXML file.

Configuration

The following parameters are recognized in the JSON parameter file:

overwrite_regions: remove previously existing text regions
xheight: height of character "x" in pixels used during training.
model: pixel-classifier model path. The special values __DEFAULT__ and __LEGACY__ load the bundled default model or previous default model respectively.
gpu_allow_growth: required for GPU use with some graphic cards (set to true, if you get CUDNN_INTERNAL_ERROR)
resize_height: scale down pixelclassifier output to this height before postprocessing. Independent of training / used model. (performance / quality tradeoff, defaults to 300)

Testing

There is a simple CLI test, that will run the tool on a single image from the assets repository.

make test-cli

Training

To train models for the pixel classifier, see its README