OCR-D Developer Guide

A practical guide to the OCR-D framework

Introduction

The “OCR-D guide” helps developers writing software and using tools within the OCR-D ecosystem.

Scope and purpose of the OCR-D guide

The OCR-D guide is a collection of concise recipes that provide pragmatic advise on how to

Notation

Lines in code examples

Words in ALL CAPS with a prepended $ are variable names:

When referring to a “something command”, it is actually ocrd something on the command line.

Other OCR-D documentation

Bootstrapping

Ubuntu Linux

OCR-D development is targeted towards Ubuntu Linux >= 18.04 since it is free, widely used and well-documented.

Most of the setup will be the same for other Debian-based Linuxes and older Ubuntu versions. You might run into problems with outdated system packages though.

In particular, it can be tricky at times to install tesseract at the right version. Try alex-p’s PPA or build tesseract from source.

Essential system packages

sudo apt install \
  git \
  build-essential \
  python python-pip \
  python3 python3-pip

Python API and CLI

The OCR-D toolkit is based on a Python API that you can reuse if you are developing software in Python.

This API is exposed via a command line tool ocrd. This CLI offers much of the same functionality of the API without the need to write Python code and can be readily integrated into shell scripts and external command callouts in your code.

So, If you do not intend to code in Python or want to wrap existing/legacy tools, a major part of the functionality of the API is available as a command line tool ocrd.

Python setup

Create virtualenv

We strongly recommend using virtualenv (or similar tools if they are more familiar to you) over system-wide installation of python packages. It reduces the amount of pain supporting multiple Python versions and allows you to test your software in various configurations while you develop it, spinning up and tearing down environments as necessary.

sudo apt install \
  python3-virtualenv \
  python-virtualenv # If you require Python2 compat

Create a virtualenv in an easy to remember or easy-to-search-shell-history-for location:

$ virtualenv -p python3.6 $HOME/ocrd-venv3
$ virtualenv -p python2.7 $HOME/ocrd-venv2 # If you require Python2 compat

Activate virtualenv

You need to activate this virtual environment whenever you open a new terminal:

$ source $HOME/ocrd-venv3/bin/activate

If you tend to forget sourcing the script before working on your code, add source $HOME/ocrd-venv3 to the end of your .bashrc/.zshrc file and log out and back in.

Install ocrd in virtualenv from pypi

Make sure, the virtualenv is activated and install ocrd with pip:

$ pip install ocrd

Generic setup

In this variant, you still need to install the ocrd Python package. But since it’s only used for its CLI (and as a dependency for Python-based OCR-D software), you can install it system-wide:

$ sudo pip install ocrd

Setup from source

If you want to build the ocrd package from source to stay up-to-date on unreleased changes or to contribute code, you can clone the repository and build from source:

$ git clone https://github.com/OCR-D/core
$ cd core

If you are using the python setup:

$ pip install -r requirements.txt
$ pip install -e .

If you are using the generic setup:

$ sudo pip install -r requirements.txt
$ sudo pip install .

Verify setup

After setting up, check that these commands do not throw errors and have the minimum version:

$ git --version
# Version 1.7 or higher?

$ make --version
# Version 9.0.1 or higher?

$ ocrd --version
# ocrd, version 0.4.0

Anatomy of an OCR-D module project (MP)

MP are git repositories with at least a description of the MP and its provided tools (ocrd-tool.json and a Makefile for installing the MP into a suitable OS.

ocrd-tool.json

This is a JSON file that describes the software of a particular MP. It serves mainly three purposes:

  1. providing a machine-actionable description of MP and the bundled tools and their parameters
  2. concise human-targeted descriptions as the foundation for the application documentation
  3. ensuring compatible definitions and interfaces, which is essential for sustainable, scalable workflows

This document is mainly focusing on the first point.

The structure and syntax of the ocrd-tool.json is defined by a JSON Schema and expects JSON Schema for the parameter definitions. In addition to the schema, the ocrd command line tool can help you validate the ocrd-tool.json

Mechanics of the ocrd-tool.json

:fire: TODO :fire:

[kba] Wir brauchen einen besseren Namen, ich kann das schon nicht mehr schreiben dauernd, ocrd-tool.json. Vielleicht einfach manifest.json oder package.json oder tool-desc odr irgendwas.

:fire: TODO :fire:

The ocrd-tool.json has two conceptual levels:

Beyond the ocrd-tool.json file, it is part of the requirements that the tools can provide the section of the ocrd-tool.json about ‘themselves’ at runtime with the -J/--dump-json flags.

The reason for this redundancy is to make the tools inspectable at runtime and to prevent “feature drift” where the software evolves to the point where the description/documentation is out-of-date with the actual implementation.

From a developer’s perspective, the easiest way to handle this is by bundling the ocrd-tool.json into your software, e.g. by the following pattern:

  1. Store the ocrd-tool.json at a location where it is easy to deploy and access after installation
  2. Symlink it to the root of the repository: ln -sr src/ocrd-tool.json .
  3. Handle --dump-json by parsing the ocrd-tool.json and sending out the relevant section
  4. Validate input and provide defaults based on the JSON schema mechanics

Metadata about the module project

Required properties are bold.

Example:

{
  "version": "0.0.1",
  "name": "ocrd-blockissifier",
  "synopsis": "Tools for reasoning about how these blocks fit on this here page",
  "git_url": "https://githbub.com/johndoe/ocrd_blocksifier",
  "dockerhub": "https://hub.docker.com/r/johndoe/ocrd_blocksifier",
  "authors": [{
    "name": "John Doe",
    "email": "johndoe@ocr-corp.com",
    "url": "johndoe.github.io"
  }],
  "bugs": {
    "url": "https://github.com/sindresorhus/temp-dir/issues"
  },
  "tools": {
      /* see next section */
    }

}

Metadata about the tools

The tools section is an object with the key being the name of the executable described and the value being an object with the following properties (bold means required):

Metadata about parameters

Required properties are bold.

required: true and setting default are mutually exclusive.

Makefile

All MP should provide a Makefile with at least two targets: deps and install.

make deps should install any dependencies, such as required python modules.

make install should install the executable(s) into $(PREFIX)/bin.

make test should start the unit/regression test suite if provided.

Makefile for Python MP

make deps should install dependencies with pip.

make install should call python setup.py install.

See the makefile of the ocrd_kraken project for an example.

Makefile for generic MP

make deps should install dependencies either by compiling from source or using apt-get.

make install should

See the makefile of the ocrd_olena project for an example.

ocrd workspace - Working with METS

METS is the container format of choice for OCR-D because it is widely used in digitzation workflows in cultural heritage institutions.

A METS file references files in file groups and can contain a variety of metadata, the details can be found in the specs.

From METS to Workspace

Within the OCR-D toolkit, we use the term “workspace”, a folder containing a file mets.xml and any number of the files referenced by the METS.

One can think of the mets.xml as the MANIFEST of a JAR or the .git folder of a git repository.

The workspace command of the ocrd tool allows various manipulations of workspaces and therefore METS files.

Git similarity intended

The workspace command’s syntax and mechanics are strongly inspired by git so if you know git, this should be familiar.

git ocrd workspace
init init
clone clone
add add
ls-files find
fetch find --download
archive pack

Set the workspace to work on

For most commands, workspace assumes the workspace is the current working directory. If you want to use a different directory, use the -d / --directory option

# Listing files in the workspace at $PWD
$ ocrd workspace find

# Listing files in the workspace at $WORKSPACE_DIR
$ ocrd workspace -d $WORKSPACE_DIR find

Use another name than mets.xml

According to convention, the METS of a workspace is named mets.xml.

To select a different basename for that file, use the -M / --mets-basename option:

# Assume this workspace structure
$ find $WORKSPACE_DIR
$WORKSPACE_DIR
$WORKSPACE_DIR/mets3000.xml

# This will fail in a loud and unpleasant manner
$ ocrd workspace -d $WORKSPACE_DIR find

# This will not
$ ocrd workspace -d $WORKSPACE_DIR -m mets3000.xml find

Creating an empty workspace

To create an empty workspace to which you can add files, use the workspace init command

$ ocrd workspace -d ws1 init
/home/ocr/ws1

Load an existing METS as a workspace

To create a workspace and save a METS file, use the workspace clone command:

$ ocrd workspace -d new-workspace clone $METS_URL
/home/ocr/new-workspace

$ find new-workspace
new-workspace
new-workspace/mets.xml

Load an existing METS and referenced files as a workspace

To not only clone the METS but also download the contained files, use workspace clone with the --download flag:

$ ocrd workspace -d $WORKSPACE_DIR clone --download $METS_URL

$ find $WORKSPACE_DIR
$WORKSPACE_DIR
$WORKSPACE_DIR/mets.xml
$WORKSPACE_DIR/OCR-D-GT-ALTO
$WORKSPACE_DIR/OCR-D-GT-ALTO/kant_aufklaerung_1784_0020.xml
$WORKSPACE_DIR/OCR-D-GT-PAGE
$WORKSPACE_DIR/OCR-D-GT-PAGE/kant_aufklaerung_1784_0020.xml
$WORKSPACE_DIR/OCR-D-IMG
$WORKSPACE_DIR/OCR-D-IMG/kant_aufklaerung_1784_0020.tif

NOTE: This will download all files, which can mean hundreds of high-resolution images. If you want more fine-grained control, clone the bare workspace and then use the workspace find command with the download flag

Searching the files in a METS

You can search the files in a METS file with the workspace find command.

See ocrd workspace --find for the full range of selection options

Downloading/Copying files to the workspace

To download remote or copy local files referenced in the mets.xml to the workspace, append the --download flag to the workspace find command:

# Clone Bare workspace:
$ ocrd workspace clone $METS_URL

$ find $WORKSPACE_DIR
$WORKSPACE_DIR
$WORKSPACE_DIR/mets.xml

# Download all files in the `OCR-D-IMG` file group
$ ocrd workspace -d $WORKSPACE_DIR find --file-grp OCR-D-IMG --download
[...]

$ find $WORKSPACE_DIR
$WORKSPACE_DIR
$WORKSPACE_DIR/mets.xml
$WORKSPACE_DIR/OCR-D-IMG
$WORKSPACE_DIR/OCR-D-IMG/kant_aufklaerung_1784_0020.tif

The convention is that files will be downloaded to $WORKSPACE_DIR/$FILE_GROUP/$BASENAME where

NOTE Downloading a file not only copies the file to the $WORKSPACE_DIR but also changes the URL of the file from its original to the absolute file path of the downloaded file.

Adding files to the workspace

When running a module project, new files are created (PAGE XML, images …). To register these new files, they need to be added to the mets.xml as a mets:file with a mets:FLocat within a mets:fileGrp, each with the right attributes. The workspace add command makes this possible:

$ ocrd workspace -d $WORKSPACE_DIR find -k local_filename
$WORKSPACE_DIR/OCR-D-IMG/page0013.tif

$ ocrd workspace -d $WORKSPACE_DIR add \
  --file-grp OCR-D-IMG-BIN \
  --file-id PAGE-0013-BIN \
  --mimetype image/png \
  --group-id PAGE-0013 \
  page0013binarized.png

$ ocrd workspace -d $WORKSPACE_DIR find -k local_filename
$WORKSPACE_DIR/OCR-D-IMG/page0013.tif
$WORKSPACE_DIR/OCR-D-IMG-BIN/page0013binarized.tif

Validating OCR-D compliant METS

To ensure a METS file and the workspace it describes adheres to the OCR-D specs, use the workspace validate command:

# Create a bare workspace
ocrd workspace -d  $WORKSPACE_DIR init

# Validate
<report valid="false">
  <error>METS has no unique identifier</error>
  <error>No files</error>
</report>

# Oops, let's set the identifier ...
$ ocrd workspace -d $WORKSPACE_DIR set-id 'scheme://my/identifier/syntax/kant_aufklaerung_1784'

# ... and add a file
$ ocrd workspace -d $WORKSPACE_DIR add -G OCR-D-IMG-BIN -i PAGE-0013-BIN -m image/png -g PAGE-0013 page0013binarized.png

# Validate again
<report valid="true">
</report>

ocrd tool – Working with ocrd-tool.json

This command helps you to explore and validate the information in any ocrd-tool.json.

The syntax is ocrd ocrd-tool /path/to/ocrd-tool.json SUBCOMMAND

ocrd-tool validate

Validate that an ocrd-tool.json is syntactically valid and adheres to the schema.

This is useful while developing to make sure there are no typos and all required properties are set.

$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json validate
<report valid="false">
  <error>[tools.ocrd-wip-xyzzy] 'steps' is a required property</error>
  <error>[tools.ocrd-wip-xyzzy] 'categories' is a required property</error>
  <error>[] 'version' is a required property</error>
</report>

This example shows that the ocrd-wip-xyzzy executable is missing the required steps and categories properties and the root level object is missing the version property.

Adding them should result in

$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json validate
<report valid="true">
</report>

Introspect an ocrd-tool.json

These commands are used for enumerating the executables contained in an ocrd-tool.json and get root level metadata, such as the version.

$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json version
0.0.1

# Lists all the tools (executables) one per-line
$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json list-tools
ocrd-wip-xyzzy
ocrd-wip-frobozz

Introspect individual tools

This set of commands allows introspection of the metadata on individual tools within an ocrd-tool.json.

The syntax is ocrd ocrd-tool /path/to/ocrd-tool.json tool EXECUTABLE SUBCOMMAND

$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json tool ocrd-wip-xyzzy dump
{
  "description": "Nothing happens",
  "categories": ["Text recognition and optimization", "Arcane Magic"],
  "steps": ["recognition/text-recognition"],
  "executable": "ocrd-wip-xyzzy"
}

# Description
$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json tool ocrd-wip-xyzzy description
Nothing happens

# List categories one per line
$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json tool ocrd-wip-xyzzy categories
Text recognition and optimization
Arcane Magic

# List steps one per line
$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json tool ocrd-wip-xyzzy steps
recognition/text-recognition

Parse parameters

The details of how a tool is configured at run-time are determined by parameters. When a parameter file is passed to a tool, it should:

The ocrd ocrd-tool tool parse-params command does just that and can output the resulting default-enriched parameter as either JSON or as shell script assignments to evaluate:

# Get JSON
$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json tool ocrd-wip-xyzzy parse-params --json -p <(echo '{"val1": 42, "val2": false}')
{
  "val1": 42,
  "val2": false,
  "val-with-default": 23
}

# Get back shell assignments to an associative array "params"
$ ocrd ocrd-tool /path/to/ocrd_wip/ocrd-tool.json tool ocrd-wip-xyzzy parse-params -p <(echo '{"val1": 42, "val2": false}')
params["val1"]="42"
params["val2"]="true"
params["val-with-default"]="23"

ocrd process - Run a multi-step workflow

OCR requires multiple steps, such as binarization, layout recognition, text recognition etc. These steps are implemented with command line tools that adhere to the same command line interface which makes it straightforward to chain these calls.

For example, to run kraken binarization and tesseract block segmentation, one could execute:

ocrd-kraken-binarize -l DEBUG -I OCR-D-IMG -O OCR-D-IMG-BIN
ocrd-tesserocrd-segment-block -l DEBUG -I OCR-D-IMG-BIN -O OCR-D-SEG-BLOCK -p tesseract-params.json

The disadvantage of individual calls is that it requires the user to check whether runs were actually successful. To remedy this, users can use the ocrd process CLI which

The same calls mentioned before can be passed to ocrd process as follows:

ocrd process -l DEBUG \
  "kraken-binarize -l DEBUG -I OCR-D-IMG -O OCR-D-IMG-BIN" \
  "tesserocrd-segment-block -l DEBUG -I OCR-D-IMG-BIN -O OCR-D-SEG-BLOCK -p tesseract-params.json"

Wrapping a CLI using bash

This section describes how you can make an existing tool OCR-D compliant, i.e. provide a CLI which implements all the specs and calls out to another executable.

For this purpose, the ocrd offers a bash library that handles:

The shell library is bundled with the ocrd command line tool and can be accessed with the ocrd bashlib command.

ocrd bashlib

To get the filename of the shell lib, use ocrd bashlib filename, which you can employ to source the shell code in a wrapper script. After sourcing this script you will have access to a number of shell functions that begin with ocrd__.

The only function you definitely need is ocrd__wrap which parses an ocrd-tool.json and scaffolds a spec-compliant CLI, parses command line arguments and parameters and lets the developer then react to the inputs.

In combination with the ocrd workspace command this allows you to write CLI applications without touching any METS or PAGE/XML files by hand.

ocrd__wrap

ocrd__wrap has this signature:

ocrd__wrap OCRD_TOOL_JSON EXECUTABLE_NAME ...ARGS

where

Example:

   ocrd__wrap /usr/share/ocrd-wip/ocrd-tool.json ocrd-wip-xyzzy "$@"