Inpainting - Based Image Compression

Inpainting-based image compression is an emerging paradigm for compressing visual data in a completely different way than popular transform-based methods such as JPEG. The underlying idea sounds very simple: One stores only a small, carefully selected subset of the data, which results in a substantial reduction of the file size. In the decoding phase, one interpolates the missing data by means of a suitable inpainting process. It propagates information from the known data into the areas where nothing has been stored, e.g. by solving a partial differential equation or by clever copy-and-paste mechanisms. Inpainting-based codecs (coders and decoders) are more intuitive than transform-based ones, they are closer to biological mechanisms in our brain, and first results show that they may offer promising performance for high compression rates. However, before these ideas become practically viable, a number of difficult fundamental problems must be solved first. They involve e.g. the selection of the data and the inpainting operator, coding strategies, and the search for highly efficient numerical algorithms. This requires a collaborative effort of experts in data compression, inpainting, optimisation, approximation theory, numerical algorithms, and biological vision. In this Dagstuhl seminar we have brought together leading researcher from all these fields for the first time. It enabled a very fruitful and inspiring interaction which will form the basis for future progress. Seminar November 13–18, 2016 – http://www.dagstuhl.de/16462 1998 ACM Subject Classification E.4 Coding and Information Theory, F.2.1 Numerical Algorithms and Problems, G.1.1 Interpolation, G.1.2 Approximation, G.1.6 Optimization, G.1.8 Partial Differential Equations, G.2 Discrete Mathematics, I.4.2 Compression (Coding)