DPE.md

Short introduction

Training to mimic the professional PS skills

Main contributions

• Augment U-Net with global features
• Improve WGAN with an adaptive weighting scheme
• Propose individual batch normalization layers for generators (better adjusting to distributions of different domains)

Architecture

Two-way GAN

contains a forward mapping and backward mapping and check the consistency.

Gen and Disc

Loss

• L2
• Loss function: Gen Loss:

Cycle consistency loss:

Adversarial losses:

Gradient penalty P:

Training strategy

Experiments

• Dataset: MIT-Adobe 5K 2250 image with retouched version for supervison training The other 2250 retouched image for target domain, untouched image of the first partition used as source domain
• Evaluation metric:
• Patchsie: 512 × 512

Final summary

Pros:

• Global feature on Unet
• Two-way GAN with individual BN

Cons:

Tips:

• WGAN uses the earth mover distance to measure the distance between the data distribution and the model distribution and significantly improves training stability.
• Instead of weight clipping, panalize the norm of the gradient of the discrimiator with respect to the input
• Detail comparision between different GAN models. Detailing paramenter tuning instructions