OCRD

A repository for online OCRD training infrastructure.

Setup

sudo apt-get install redis
pip install -r requirements.txt

Installation of the Engines

The instructions for installing the engines (Tesseract, kraken, OCRopus, Calamari) are in the folder install.

Installaton of the Training System

$ wget https://github.com/Doreenruirui/okralact/archive/master.zip

Deploy the system on a server

Coming soon.

System Structure

Okralact is both a set of specifications and a prototype implementation for harmonizing the input data, parameterization and provenance tracking of training different OCR engines. It is a client/server architecture application. The interactions between the client nodes and the server are implementeqd using Flask, a lightweight web application framework for Python. All the training or evaluation jobs submitted to the server are handled in the background by task queues implemented wth Redis Queue (RQ).

The folder app contains the code for the web-application and task queues. The folder app/templates defined a set of web interfaces that the users directly interacted with. Through these webpages, the user can

upload
- a dataset
- a configuration file for training a specific engine
submit
- a job for training or fine tuning an OCR model by selecting a uploaded model and a configuration file
- a job for generating the report of validating all the intermediately models saved during training (can be used to pick the best model)
- a job for evaluating a trained model by selecting a evaluation dataset
download
- trained model
- validation report, where the character error rate and word error rate for each itermediate model could be found. The frequency to save the intermediate is defined by the parameter "save_frequency". User can use the report to decide which model to use and see the training curve.
- evaluation report, contains the evaluation result for a user selected model on a user selected dataset.

The file app/route.py contains the code for handling the users' HTTP requests and responses with Flask, and the code for processing the users' jobs with Redis Queue. There are three task queues running in the backend, i.e., the training queue, the validation queue and the evaluation queue. Each type of job submitted by the user would be put into its corresponding task queue, where it will wait to be excuted.

When a job gets opportunity to run, it will call the common API for training, validation or evaluation defined in the folder engines. The jobs in training task queue, validating task queue and evaluating task queue will call the API defined in the file train.py, valid.py and eval.py respectively.

The unifie specifications for all the OCR engines are defined in the folder engines/schemas with jsonschema. It contains the following files:

common.schema , defines a set of unifed parameters for each engine.
engine_[engine name].schema, defines the parameter set for each engine.
engines/schemas/models, contains the schemas for the neural network structure for all the engines:
- model_[engine name].schema, define the neural network layers, e.g., input, cnn, pooling, dropout, rnn, output allowed by each engine.
- layer_all_[engine name].schema, define the parameters allowed in each layer by each engine.
- layer_[layer name].schema, defines the set of parameters allowed by each neural network layer.
translate.json, defines the mapping of the unified parameters to the engine-specific parameters.
validating parameters.py, validate whether the configuration file that the users uploaded are valid.
translate_parameters.py, translate the user-specific configuration files into the original set parameters that can recognized by each engine.
translate_model.py, translate the unified definition of neural network structures to the engine -specific model structure definition language.
process_tesseract.py, define the steps to preprocess the data fo tesseract.

The folder evaluation contains the files for evaluating the OCR results by comparing them to the ground truth. It is based on the evaluation code from OCRopus, where we add the word error rate in additional to the character error rate and confusion matrix.