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.
- 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.)
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.