Repository files navigation

Cleaning printed text using Denoising Autoencoder based on UNet architecture in PyTorch

The UNet architecture used here is borrowed fromhttps://github.com/jvanvugt/pytorch-unet.The only modification made in the UNet architecture mentioned in the above link is the addition of dropout layers.

• torch >= 0.4
• torchvision >= 0.2.2
• opencv-python
• numpy >= 1.7.3
• matplotlib
• tqdm

Set the number of total synthetic images to be generatednum_synthetic_imgs and set the percentage of training datatrain_percentage inconfig.pyThen run

python generate_synthetic_dataset.py

It will generate the synthetic data in a directory nameddata (can be changed in the config.py) in the root dirctory.

Set the desired values oflr,epochs andbatch_size inconfig.py

Inconfig.py,

• setresume to False

python train.py

Inconfig.py,

• setresume to True and
• setckpt to the path of the model to be loaded, i.e. ckpt = 'model02.pth'

python train.py

The model was trained for 12 epochs for the configuration mentioned inconfig.py

Inconfig.py,

• setckpt to the path of the model to be loaded, i.e. ckpt = 'model02.pth'
• settest_dir to the path that contains the noisy images that you need to denoise ('data/val/noisy' by default)
• settest_bs to the desired batch size for the test set (1 by default)

python test.py

Once the testing is done, the results will be saved in a directory namedresults

*

*

*

*

*