Autonomous detection of information patterns through hierarchical peeling

Abstract

Finding useful information in real world scenes is very important for many scene understanding tasks. Signs, scripts, information panels, and logos typically stand out from their environment for a human observer, but algorithmically locating them seems to depend on the actual context. We propose a hierarchical method that locates potentially interesting areas and examines whether their content is worth further examination. The algorithm digs into the containers by iteratively peeling down external shells and deciding what parts constitute the foreground. The proposed algorithm can be efficiently implemented on topographic processor arrays, since it mostly contains standard topographic instructions. The effectiveness of the approach is demonstrated through several examples.