2019 CVPR Unsupervised Domain-Specific Deblurring via Disentangled Representations阅读笔记

Abstract

1. The disentanglement is achieved by splitting the content and blur features in a blurred image using content encoders and blur encoders.
2. 4 kind of loss functions play different role

Qs:

1. what is content? what is blur features? How encoder them?
2. How these loss functions work?

Introduction

1. Priors based method could not generalize to specific image domains, like face, text and low-illunination images.
2. Nimisha[25] proposed an unsupervised image deblurring method based on GANs where they add reblur loss and multi-scale gradient loss on the model. Their results are on some real blurred images are not satisfactory.
3. Cycle-GAN, and DualGAN often encode other factors rather than blur information into the generators.

Network