peijl1998 / Modification-on-DPED

A toy experiment for modifying DPED

Toy Experiment for modifying DPED

1. What is DPED?

DPED is "DSLR-Quality Photos on Mobile Devices with Deep Convolutional Networks".

• Paper:[paper] [project webpage]
• Repo:DPED

In a word, it turns a poor phone-quality photo to a DSLR-quality photo with GAN network and other technologies.

2. What I've tried to modify?

• The comparison between batch-normalization and instance-normalization.
• Two generator: Densenet and Resnet.
• Paired training to Unpaired training based on Cycle GAN network.(By introducing cycle loss.)

3. Some results

Tips: For the limitation of time and ability, we doesn't focus on the results(two index: psnr&ssim). This is just a toy experiment. :)

More details are recored in the report

Keywords: GAN, Cycle GAN, Densenet, Resnet, Batch-norm, Instance-norm, Tensorflow.

1. Loss

   

2. Visual improvement(iteration 1000 and iteration 19000)