Pattern-oriented algorithmic complexity: towards compression-based information retrieval

Kurzfassung

The assimilation of informational content to computational complexity is more than 50 years old, but a way of exploiting practically this idea came only recently with the definition of compression-based similarity measures, which estimate the amount of shared information between any two objects. These techniques are effectively employed in applications on diverse data types with a universal and basically parameter-free approach; nevertheless, the difficulties in applying them to large datasets have been seldom addressed. This thesis proposes a novel similarity measure based on compression with dictionaries which is faster compared to known solutions, with no degradations in performance; this increases the applicability of these notions, allowing testing them on datasets with size up to 100 times larger than the ones previously analyzed in literature. These results have been achieved by studying how the classical coding theory in relation with data compression and the Kolmogorov notion of complexity allows decomposing the objects in an elementary source alphabet captured in a dictionary, regarded as a set of rules to generate a code having semantic meaning for the image structures: the extracted dictionaries describe the data regularities, and are compared to estimate the shared information between any two objects. This allows defining a content-based image retrieval system which requires minimum supervision on the user’s side, since it skips typical feature extraction steps, often parameter-dependant; this avoids relying on subjective assumptions which may bias the analysis, adopting instead a data-driven, parameter-free approach. Applications are presented where these methods are employed with no changes in settings to different kinds of images, from digital photographs to infrared and Earth Observation (EO) images, and to other data types, from texts and DNA genomes to seismic signals.