Project IV 2014-2015

Brainy image compression

Description

While the raw data coming out of today's digital camera sensors can easily reach tens of megabytes per image, compression formats such as JPEG bring this down to a more manageable size. The trade-off is of course quality loss, and finding a good balance relies on our understanding of which features of an image are important for the human eye.

Traditional compression formats are hand-engineered and often do not work well for particular types of images. An alternative approach, known under the name of auto-encoding, tries to use ideas from neural networks to construct compression algorithms that tune themselves to the class of images that they are asked to compress. They are based on statistical methods and concepts from condensed matter physics, and thus quite different from the old-style compression methods.

In this project you will learn the maths behind old and new image compression mechanisms. You will get to experiment with them on the computer, and perhaps even build your own ideas on top.

Prerequisites

Some affinity with statistical mechanics is useful, as is the willingness to play with ideas on the computer (you will be given a base of Python examples to build on).

Resources and references