abadredd / docproj

Document Rectification and Illumination Correction using a Patch-based CNN

The Flatteners

Project Page

Deep Learning U-Net Model Reference

DocProj Project Page

This work was heavily influenced by the DocProj project. Please checkout their code at the link above.

Dependencies

• Python3
• Windows (and Linux for Dataset Generation)
• CUDA and CuDNN

Generate Your Own Dataset

Generating your own dataset is a great way to train an existing model. Please follow the instructiosn on the main page for dataset generation.

Generate Data-Set Patches

Please run the following pre-processing commands to create local and global patches of your data-set. These patches are necessary inputs for model training. Please change the arguments as necessary for your implementation.

python local_patch.py

python global_patch.py

Training

Please run the follwong command for training, and change the arguments as necessary for your implementation.

python train.py

Evaluate an Image with a Model

Please run the following four command to evaluate the model. An example is given in training.sh. Additionally, the Graphcut.exe application can be found here.

python eval.py --imgPath [input_image.png] --modelPath [model_to_save.pkl] --saveImgPath [new_resized_image.png] --saveFlowPath [myflow.npy]

Graphcut.exe [myflow.npy] [my_new_flow.npy]

python resampling.py --img_path [new_resized_image.png] --flow_path [my_new_flow.npy]

python.exe eval_illumination.py --imgPath [resamplling_result.png] --savPath [output.png] --modelPath [model_illNet.pkl]

Contact

Arnav Sharma - [email protected]

Austin Jeffries - [email protected]

Ali Badreddine - [email protected]