Skip to content

Latest commit

 

History

History
31 lines (23 loc) · 1.04 KB

readme.md

File metadata and controls

31 lines (23 loc) · 1.04 KB

SAR2SAR

This projet is an implementation in pytorch of SAR2SAR, a denoised deep neural network for SAR images denoising. The network was trained on TerraSAR-X images

To run this code you must have :

  • Pytorch
  • numpy
  • opencv
  • matplotlib

Predictions

To make prediction with this network you have to :

  • put .npy images in 'data/real'
  • in a cmd type 'python predict.py'
  • results will be stored in 'data/results/real'

If you want to add artificial speckle on your images you have to set add_speck to True line 19 of predict.py

Training

To train the network you have to :

  • put phase A .npy train images in data/train_A
  • put phase A .npy eval images in data/eval
  • put phase B & C train images in data/train_BC. Each image pile must be in a different folder, if you have only one pile create a single folder
  • put phase B & C eval images in data/eval_real
  • in a cmd type 'python train_ABC.py'
  • network weights will be stored in 'pipeline/out/unsupervised'
  • in data/sample you will have a denoised version of eval images for each epoch