eduzhai > Applied Sciences > Engineering >

Image Deconvolution via Noise-Tolerant Self-Supervised Inversion

  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: We propose a general framework for solving inverse problems in the presenceof noise that requires no signal prior, no noise estimate, and no cleantraining data. We only require that the forward model be available and that thenoise be statistically independent across measurement dimensions. We build uponthe theory of $ mathcal{J}$-invariant functions (Batson & Royer 2019,arXiv:1901.11365) and show how self-supervised denoising emph{à la}Noise2Self is a special case of learning a noise-tolerant pseudo-inverse of theidentity. We demonstrate our approach by showing how a convolutional neuralnetwork can be taught in a self-supervised manner to deconvolve images andsurpass in image quality classical inversion schemes such as Lucy-Richardsondeconvolution.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×