Source codes of our CS591 project: Arabic Document Analysis Using Deep Learning Structure.
Wenda Qin, Hao Yu
A pre-trained model can be downloaded here: http://cs-people.bu.edu/wdqin/current_model_nms_05
- training the model: run Faster-RCNN_based_and_CRAFT_based_processing/creating_masks_for_BCE.ipynb then run Faster-RCNN_based_and_CRAFT_based_processing/training_detection_model.ipynb you will need to modify the dataset size in the code as the default number of images in the code is 300.
- to watch testing results generated by the model: run Faster-RCNN_based_and_CRAFT_based_processing/Visualizing_the_model.ipynb, you need to specify the model you trained or the pre-trained one. run Faster-RCNN_based_and_CRAFT_based_processing/Conversion_of_Format.ipynb the visualized results are shown in Faster-RCNN_based_and_CRAFT_based_processing/results (you might need to change the path correctly before doing that)
The implementation of CRAFT-based method are in the folder craft and segmentation
- PyTorch>=0.4.1
- torchvision>=0.2.1
- opencv-python>=3.4.2
- requirments.txt
pip install -r craft/requirements.txt
-
Download the trained model which is trained on SynthText, IC13, IC17 dataset.
-
Run with pretrained model
python craft/test.py --trained_model=[weightfile] --test_folder=[folder path to test images]
The result images will be saved to ./result by default.
- Train
python segmentation/main.py --train --input=[folder path to input images] --output=[folder path to output files]
- Test
python segmentation/main.py --input=[folder path to input images] --output=[folder path to output files]
- Merging non-text results from Faster-R-CNN-based method: run CRAFT_generate_csv.ipynb run Merging_Text_and_non_text_csvs.ipynb
To visualize the result, run Faster-RCNN_based_and_CRAFT_based_processing/Conversion_of_Format.ipynb the visualized results are shown in Faster-RCNN_based_and_CRAFT_based_processing/results (you might need to change the path correctly before doing that)