Tradeoffs in restoration problems

הרקע לפרויקט:

Restoration algorithms are oftentimes evaluated by different criteria. For example, we may want a restored image to be both accurate (low distortion) and perceptually pleasing (as if it is a natural image). A seminal paper have shown that there is a tradeoff between these properties. In this project we will study extensions of this phenomenon. Primarily, we will explore tradeoffs between low-level tasks (e.g., restoring an image with low distortion) and high-level tasks (classifying the restored image).

מטרת הפרויקט:

The goal of this project is to identify and explore (empirically and theoretically) tradeoffs between low-level restoration and high-level tasks.

תכולת הפרויקט:

1. Understanding known tradeoffs (e.g., perception-distortion) and techniques to explore them (convex optimization, rate-distortion theory).
2. Examining if they can be sharpened.
3. Identifying tradeoffs between low-level restoration and high-level tasks (e.g., classification).
4. Empirical and theoretical analysis.

קורסי קדם:

קורס מבוא ללמידת מכונה, עדיפות לרישום לקורסים שערוך פרמטרים ולמידה עמוקה

מקורות:

https://arxiv.org/abs/1711.06077