Skip to main content

Image Denoising using Unet

Created on April 1|Last edited on April 1

1020304050epoch0.60.650.70.750.80.85
1020304050epoch182022242628
1020304050epoch0.0020.0030.0040.0050.0060.0070.0080.0090.010.020.03
1020304050epoch0.50.60.70.8
1020304050epoch182022242628
1020304050epoch0.0020.0030.0040.0050.0060.0070.0080.0090.01
Run set
5

Best Hyperparameters:
  • Channel Dimensions of Encoder/Decoder(channel=2x) : 128, 256, 512, 1024 performed better than 64, 128, 256,512 (channel=1x)
  • Runs with (channel=2x) converge in 25-30 epoochs
  • Training with batch size = 128 is more stable than bs= 64.
  • Tried runs with two types of Residual Connections: Additive Based and standard Concatenation Based
Best Run: Run6: with hyperparameters:
  • Channel Dimensions of Encoder/Decoder(channel=2x) : 128, 256, 512, 1024
  • Residual Connection of Concatenation type.
  • Batch Size = 128
  • Epochs = 27
  • Patch Size = 50, stride = 25