Learning to denoise without clean examples

הרקע לפרויקט:

In this project we will explore methods to train a denoiser from unaligned noisy images. The deep learning approaches that have been proposed in the past years for training a denoiser without g.t. clean data do not exploit connection between mismatched images of the same object.

מטרת הפרויקט:

The goal of the project is to design a deep learning method for training a denoiser from unaligned noisy images that outperforms existing methods (which currently do not fully the information shared between such images).

תכולת הפרויקט:

1. Exploring the performance of current methods when trained on unaligned noisy images (starting with noisy shifted versions) and identifying gaps.
2. Proposing a mathematically-backed method for the problem at hand.
3. Training a model that outperforms existing alternatives.
4. Potentially: generalizing the idea from shifts to more general transformations.

קורסי קדם:

קורס מבוא ללמידת מכונה, הרשמה לקורס למידה עמוקה

מקורות:

* https://arxiv.org/abs/1803.01314
* https://arxiv.org/abs/1803.04189