three models are used and assembled together in this implementation, there are:
- an auto encoder covnet is used to remove noise
- a traditional 2d median filter algorithm is used to remove very dark noise
- an unet is used as the 'assembling' network
- all models are pixel-wise models, so there is no resize operation. In prediction, the target image is cropped into small blocks which fit the network input.
left column shows the target images, ground truth images are in the middle, the right column shows the prediction of the model