Keras-ocr-cpp

Keras OCR implemented in C++ based on faustomorales/keras-ocr repo. This is a small side project for my research, where python couldn't be used. Thus, it may lack some of the features available in original repo.

---

Getting started

Requirements

In this project two libraries were used:

• cppflow - C++ wrapper for tensor manipulation and eager execution. Installation is described in repo.
• Tensorflow C API - use their documentation or description in cppflow docs to install TF C API locally/globally. For Apple Silicon CPUs compilation from sources is required, because the version provided in docs is only for x86.

Models

After installing the dependencies, export the models to protobuf format (model + weights). This can be done using python in keras-ocr. It's best to use Poetry - python dependency manager. Follow the documentation to install poetry.

Install all requirements for keras-ocr, activate created venv and run python.

poetry install
poetry shell
python

Then, execute following code, which will export models to specified directories.

import keras_ocr


pipeline = keras_ocr.pipeline.Pipeline()
pipeline.detector.model.save('/path/to/models/craft-detection')
pipeline.recognizer.prediction_model.save('/path/to/models/crnn-recognition')

Export or move them to models directory or modify CMakeLists.txt.

Usage

mkdir build
cd build
cmake ../keras_ocr
cmake --build .
./KerasOCR

Example images were downloaded from Wikipedia. Sources:

• img1.jpg: https://upload.wikimedia.org/wikipedia/commons/b/bd/Army_Reserves_Recruitment_Banner_MOD_45156284.jpg
• img2.jpg: https://upload.wikimedia.org/wikipedia/commons/b/b4/EUBanana-500x112.jpg

Advanced configuration

The execution of inference is entirely on the side of cppflow. If you need to make changes, install the library locally and make modifications to it.