DeepRule Deep Learning Framework

The original implementation uses an Azure web service. If you do not have access to Azure, you can replace the ocr_result(image_path) function using the pytesseract package.

The function must return a dictionary (or similar structure) where each word entry contains:

• text: The string version of the word.
• boundingBox: The bounding box in the format [topleft_x, topleft_y, bottomleft_x, bottomleft_y].

To use a trained model as a web server pipeline, run the following command:

python manage.py runserver 8800

Once running, you can interact with the service at localhost:8800.

DeepRule requires an Anaconda environment and a machine with a GPU and CUDA installed.

1. Create the environment using the provided package list:

conda create --name DeepRule --file DeepRule.txt

2. Activate the environment:

source activate DeepRule

If the original DeepRule.txt fails to install, try using requirements-2023.txt instead.

You must compile the C++ implementation of the corner pooling layers. This implementation is based on CornerNet. Navigate to the specific directory within your CornerNet installation and run the setup script:

cd <CornerNet dir>/models/py_utils/_cpools/
python setup.py build_ext --inplace

To train a model, you need a JSON configuration file in config/ and a corresponding .py model file in models/.

Module Configuration Names:

• Bar: CornerNetPureBar
• Pie: CornerNetPurePie
• Line: CornerNetLine
• Query: CornerNetLineClsReal
• Cls: CornerNetCls

Training Command:

python train_chart.py --cfg_file <model_name> --data_dir <data_path>
# Example:
python train_chart.py --cfg_file CornerNetBar --data_dir /home/data/bardata(1031)

Compile the NMS code (derived from Faster R-CNN and Soft-NMS) by running make in the external directory of your CornerNet installation:

cd <CornerNet dir>/external
make

Install the required MS COCO APIs using pip:

pip install pycocotools

To test a batch of images directly, you must pre-assign the chart type using the test_pipe_type_cloud.py script:

python test_pipe_type_cloud.py --image_path <image_path> --save_path <save_path> --type <type>
# Example:
python test_pipe_type_cloud.py --image_path /data/bar_test --save_path save --type Bar

The dataset contains different formats depending on the chart type:

• Pie data: bbox is [center_x, center_y, edge_1_x, edge_1_y, edge_2_x, edge_2_y], representing the three critical points for a sector.
• Line data: bbox is [d_1_x, d_1_y, ..., d_n_x, d_n_y], representing the data points for a line.
• Bar data: bbox is the standard bounding box of the bars.
• Cls data: bbox is the standard bounding box. Different category_id values refer to different components (e.g., draw area, title, legends).