13.11.16 - 18.11.16, Seminar 16462

Inpainting-Based Image Compression

Diese Seminarbeschreibung wurde vor dem Seminar auf unseren Webseiten veröffentlicht und bei der Einladung zum Seminar verwendet.

Motivation

Since the amount of visual data is rapidly increasing, there is a high demand for powerful methods for compressing digital images. A well-known example is the lossy JPEG standard that is based on the discrete cosine transform. Unfortunately ist quality deteriorates substantially for high compression rates, such that better alternatives are needed.

The goal of this seminar is to pursue a completely different strategy than traditional, transform-based codecs: We study approaches that rely on so-called inpainting methods. They store only a small, carefully selected subset of the image data. In the decoding phase, the missing data is reconstructed by interpolation with partial differential equations (PDEs) or by copying information from patches in other image regions. Such codecs allow a very intuitive interpretation, and first experiments show their advantages for high compression rates where they can beat even advanced transform-based methods.

However, inpainting-based codecs are still in an early stage and require to solve a number of challenging fundamental problems, for example:

  1. Which data gives the best reconstruction?
  2. What are the optimal inpainting operators?
  3. How should the selected data be encoded and decoded?
  4. What are the most efficient algorithms for real-time applications?
  5. Can these approaches be tailored towards specific imagery and other data types (e.g. hyperspectral data, HDR images, videos, surface data)?

These problems are highly interrelated. Moreover, they require interdisciplinary expertise from various fields such as image inpainting, data compression and coding, approximation theory, optimisation, and numerical analysis. To design these codecs in an optimal way, one must also understand their connections to other image compression strategies and to closely related areas such as biological vision, perceptually relevant features, texture models, parameter identification, harmonic analysis, sparsity and compressed sensing, scattered data approximation, radial basis functions, subdivision strategies, geometric modeling, and computer graphics.

There has been no workshop on this topic before. Thus, we want to bring together for the first time leading researchers from these different fields and with different background, ranging from computer science over mathematics and electrical engineering to psychophysics. They can benefit from their interactions in the highly inspiring atmosphere of Schloss Dagstuhl. It is also planned to edit the first book that is solely devoted to this topic, and participants are invited to contribute to it.