-
Only cats pictures from Dogs vs. Cats dataset were used for denoising because why not (^._.^)
-
All pictures were resized to 128x128 and normalized in range [-1, 1]
-
Gaussian noise with a mean=0 and variance=0.08 was applied to the images
-
10000 pictures were used for training and 2500 for testing
-
Total parameters of autoencoder: 776,579
-
The training was done for 50 epochs
-
Learning rate scheduler was used to reduce learning rate while training
-
Autoencoder architecture:
Input
Images with a shape [3, 128, 128] ->
Encoder
-> 4 blocks of [Conv2d -> LeakyReLU activation] ->
Latent space
-> MaxPool2d -> MaxUnpool2d ->
Decoder
-> 3 blocks of [ConvTranspose2d -> LeakyReLU activation] -> ConvTranspose2d -> Tanh activation ->
Output
-> denoised images with a shape [3, 128, 128]
Some denoised images. More of them you can find here
