dbetm / handwritten-flowchart-with-cnn

Recognition of handwritten flowcharts using convolutional neural networks to generate C source code and reconstructed digital flowchart.

Recognition of handwritten flowcharts with CNNs

Recognition of handwritten flowcharts using convolutional neural networks to generate C source code and reconstructed digital flowcharts.

Overview

The pipeline implemented in order to solve the problem of handwritten flowchart recognition uses image preprocessing, the input image is sent to two detectors, the shape-connector detector and the text detector. For text flow, the image is binarize and it uses Keras OCR to locate text and an implemented model with CNN + LSTM for character classifing; on the flow of shapes and connectors it uses unsharp masking and a model that is called Faster R-CNN with backbone VGG-16, which is an object detection model.

In order to augment the precision of the text detector, the technique called continual learning is used. After, some training with the text style of a specific user, the model will improve the text recognition.

Finally, the outputs are the generated source code in C, its compilation output and the digital reconstructed diagram as an image.

Note: Flowcharts used for testing are constructed with a defined shape-connector set. You can check it here.

Set up for testing detections

1. Create a virtual environment (venv) with Conda with name handwritten-flowchart-recog. The project was tested on Python 3.6 and Python 3.7. So, consider to use the same version.
2. Download / clone this repo.
3. Acivate the new venv, move to the project directory and install the requirements: $ pip install -r requirements.txt
4. Shapes-connectors model:
   • Download the folder from here - I will give you access on Google Drive as soon as possible.
   • Paste it (unzipped) into model/training_results/ (path inside the repo), so must be model/training_results/9
5. Text model:
   • Download IAM dataset from here.
     • Inside text_model, please create a folder with name data_model.
     • iam.hdf5 (94.1 MB) paste into text_model/data_model/
   • Download pre-trained model from here.
     • checkpoint_weights.hdf5 (38.5 MB) paste into text_model/output/iam/puigcerver/

Usage

1. Please, activate your Conda enviroment. $ conda activate handwritten-flowchart-recog
2. Move to inside repository folder, example: $ cd handwritten-flowchart-with-cnn
3. Type: $ python3 handler.py . Alternatively you can pass the Conda env. name: $ python3 handler.py --env another-conda-env
4. Use "Recognize flowchart" option to process a handwritten flowchart.
5. Click at "Predict" button.
6. A window with the shape detection will appear. You can close it pressing 'q' key or using the X button.
7. A window with the text detection will appear. You can close it pressing 'q' key or using the X button.
8. The final results will be saved on results/results_x.

Some examples of the results

Calculate the nth term of the Fibonacci sequence.

Hello world.

---

Paper

A paper was written in 2022 and published on International Journal of Computer Applications, you can find it here: Recognition of Handwritten Flowcharts using Convolutional Neural Networks

---

Dataset

Would you like to download the training dataset? Link to Kaggle. On Kaggle you will find details about it.

Please cite the dataset with:

• Author: ISC UPIIZ students
• Title: Flowchart 3b
• Version: 3.0
• Date: May 2020.
• Editors: Onder F. Campos and David Betancourt.
• Publisher Location: Zacatecas, Mexico.
• Electronic Retrieval Location: https://www.kaggle.com/davbetm/flowchart-3b

Notes

• This project was finished on July 2020 as a final school project, since then, only minor fixes and improvements have happened.