eduzhai > Applied Sciences > Engineering >

Variational Image Restoration Network

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 15 pages

Abstract: Deep neural networks (DNNs) have achieved significant success in imagerestoration tasks by directly learning a powerful non-linear mapping fromcorrupted images to their latent clean ones. However, there still exist twomajor limitations for these deep learning (DL)-based methods. Firstly, thenoises contained in real corrupted images are very complex, usually neglectedand largely under-estimated in most current methods. Secondly, existing DLmethods are mostly trained on one pre-assumed degradation process for all ofthe training image pairs, such as the widely used bicubic downsamplingassumption in the image super-resolution task, inevitably leading to poorgeneralization performance when the true degradation does not match with suchassumed one. To address these issues, we propose a unified generative model forthe image restoration, which elaborately configures the degradation processfrom the latent clean image to the observed corrupted one. Specifically,different from most of current methods, the pixel-wisely non-i.i.d. Gaussiandistribution, being with more flexibility, is adopted in our method to fit thecomplex real noises. Furthermore, the method is built on the general imagedegradation process, making it capable of adapting diverse degradations underone single model. Besides, we design a variational inference algorithm to learnall parameters involved in the proposed model with explicit form of objectiveloss. Specifically, beyond traditional variational methodology, two DNNs areemployed to parameterize the posteriori distributions, one to infer thedistribution of the latent clean image, and another to infer the distributionof the image noise. Extensive experiments demonstrate the superiority of theproposed method on three classical image restoration tasks, including imagedenoising, image super-resolution and JPEG image deblocking.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×