JPEG-ARTIFACT-REMOVER

JPEG artifact remover using GAN architecture

Model is trained based on NoGan approach , initally both generator and discriminator is trained separately.After that , Normal Gan method is followed for both generator and discriminator.

Model Architecture

In this work, i followed the approach proposed in this paper ,

A NoGAN approach for image and video restoration and compression artifact removal

link

I did lot of modifications as per my requirement, the intuition behind patchbased training,NoGan method is taken from the paper.

• Generator model will have double Residual in Residual Dense Block (RRDB) followed convolution layers
• Discriptor model will have 6 Convolution layer having leaky Relu as an activation function

Data Preparation

Generator Training

• 900 images is take from the Div2kdataset
• train_data = 700 , test_data = 100 , val_data = 100
• Initially patches or cropping is done on the images by using albumentations library
• JPEG compression is done with a Quality Factor = 10

Discriminator Training

• Two input image is being passed, real high resolution image and reconstructed image from jpeg.
• train_data = 300 , test_data = 50 , val_data = 50

Gan Traning

• Same data used for the generator is used
• Model was able to produce 0.67 SSIM and 24 PSNR on the test dataset and able to generalize well across all the images

Here as of resource constraints, batchsize = 1.

Losses

• Intially for the generator, there will be two losses, pixel loss (mse) and perceptual loss(VGG feature mse)
• For the Discriptor, Binary cross entropy is used
• During GanTraining, for generator ,in addition to above losses binary cross entroy is added as generator want the fake to be classified as real.