Image Denoising GAN

The task was, given a set of degraded images, degraded using some unknown function, one has to build a GAN to fix the images and bring them as closer to the ground truth as possible.

Generator Architecture

The generator is having a ResNet like architecture with skip connections of length 3. Each residual block consists of 3 convoltion blocks activated using LeakyReLU.

Discriminator Architecture

The discriminator network is having a VGG like architecture having Convolutions, Batch Norm activated with LeakyReLU

Loss Function Used

The loss function of the generator consists of a linear combination 3 losses. Adversarial component of the generator equation Pixelwise MSE with respect to the ground truth L1 Perceptual loss on features obtained from VGG16 model

Solution Implemented

Trained the model on a compound loss function consisting of pixelwise loss, and finally the adversarial loss. L1 Perceptual loss on features obtained from VGG16 model can be used for improving results. Training enabled faster with different values of learning rates of discriminator and Generator.

Evaluation Metric Used: Peak signal to Noise Ratio (PSNR)

Sample