laxris/DirtyDocuments

Scanned Document Denoiser

Implementation of an ensemble model with CNN autoencoder and image processing as base learners, and CNN or XGB meta learner, and CycleGAN for cleaning scanned documents, to faciliate OCR subsequently.

For further details of the approach, my Medium article explains the full project and results.

To use the ensemble model (CNN or XGB):

• Put dirty images in data/x_train, associated target clean images in data/y_train, dirty test images in data/x_test.
• Run data_augmentation.ipynb
• Run base_learners.ipynb
• Run stacker_models.ipynb

To use the CycleGAN model:

• Put dirty images for training/testing in gan_data/trainA / gan_data/testA
• Put clean images in gan_data/trainB / gan_data/testB
• The dirty and clean images do not need to be paired
• Run CycleGAN.ipynb

References

• Denoising Dirty Documents on Kaggle
• Image Processing in OpenCV
• Image Processing + Machine Learning in R: Denoising Dirty Documents Tutorial Series by Colin Priest
• Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, by Zhu et al.
• Keras Implementation of CycleGAN