eduzhai > Applied Sciences > Engineering >

Noise2Inpaint Learning Referenceless Denoising by Inpainting Unrolling

  • Save

... pages left unread,continue reading

Document pages: 9 pages

Abstract: Deep learning based image denoising methods have been recently popular due totheir improved performance. Traditionally, these methods are trained in asupervised manner, requiring a set of noisy input and clean target image pairs.More recently, self-supervised approaches have been proposed to learn denoisingfrom only noisy images. These methods assume that noise across pixels isstatistically independent, and the underlying image pixels show spatialcorrelations across neighborhoods. These methods rely on a masking approachthat divides the image pixels into two disjoint sets, where one is used asinput to the network while the other is used to define the loss. However, theseprevious self-supervised approaches rely on a purely data-driven regularizationneural network without explicitly taking the masking model into account. Inthis work, building on these self-supervised approaches, we introduceNoise2Inpaint (N2I), a training approach that recasts the denoising probleminto a regularized image inpainting framework. This allows us to use anobjective function, which can incorporate different statistical properties ofthe noise as needed. We use algorithm unrolling to unroll an iterativeoptimization for solving this objective function and train the unrolled networkend-to-end. The training paradigm follows the masking approach from previousworks, splitting the pixels into two disjoint sets. Importantly, one of theseis now used to impose data fidelity in the unrolled network, while the otherstill defines the loss. We demonstrate that N2I performs successful denoisingon real-world datasets, while better preserving details compared to its purelydata-driven counterpart Noise2Self.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×