This repository contains the codes for different techniques for reducing JPEG compression artifacts in images
- ARGAN folder contains the code which employs adversarial training for reducing compression artifacts. The generator network consists of a residual network with 5 residual blocks.
- The discriminator network consists of 8 convolutional layers and 2 fully connected layers (no pooling layers).
- Some of the hacks articulated in Gan hacks have been used.
- The overall architecture has been inspired from SRGAN paper
- baseline folder contains the code for several baseline experiments without using any adversarial training.
- We use a generator similar to the ARGAN folder (5 residual blocks), which aims to reduce the JPEG compression artifacts conditioned on the input compressed image.
- Variety of loss functions can be used; we report the model's performances on a) only L2 loss b) L2 + Total variation (TV) c) L2 + perceptual loss.
- Perceptual loss has inspired from this paper, for we’ve used a pre-trained VGG network and compared the activation of 4 th conv layer in the 5th block for the ground truth and the reconstructed image.
Python 2.7, Tensorflow 1.0.1
The following are reported after 100000 iterations (which takes about 2 days when run on Nvidia GeForce GTX Titan X) and tested as average on 29 images of LIVE1 dataset
-
ARGAN: Noisy PNSR = 28.06 dB | Reconstructed PNSR = 29.02 dB | Improvement = 0.96 dB
-
Baseline 1 (only L2 loss): Noisy PSNR = 28.06 dB | Reconstructed PSNR = 28.82 dB | Improvement = 0.76 dB
-
Baseline 2 (L2 + TV): Noisy PSNR = 28.06 dB | Reconstructed PSNR = 23.54 dB | Improvement = -4.52 dB
-
Baseline 3 (L2 + perceptual loss): Noisy PSNR = 28.06 dB | Reconstructed PSNR = 28.77 dB | Improvement = 0.71 dB
-
Baseline 4 (0.9xL2 + 0.1xperceptual loss): Noisy PSNR = 28.06 dB | Reconstructed PSNR = 28.83 dB | Improvement = 0.77 dB
