New paradigm of learnable computer vision algorithms based on the representational MDL principle

Abstract

Learning is one of the most crucial components, which increases generality, flexibility, and robustness of computer vision systems. At present, image analysis algorithms adopt particular machine learning methods resulting in rather superficial learning. We present a new paradigm for constructing essentially learnable image analysis algorithms. Learning is interpreted as optimization of image representations. Notion of representation is formalized within information-theoretic framework. Optimization criterion is derived from well-known minimum description length (MDL) principle. Adaptation of the MDL principle in computer vision has been receiving increasing attention. However, this principle has been applied in heuristic way. We deduced representational MDL (RMDL) principle that fills the gap between theoretical MDL principle and its practical applications. The RMDL principle gives criteria both for optimal model selection of a single image within given representation, and for optimal representation selection for an image sample. Thus, it can be used for optimization of computer vision systems functioning within specific environment. Adequacy of the RMDL principle was validated on segmentation-based representations applied to different object domains. A method for learning local features as representation optimization was also developed. This method outperformed some popular methods with predefined representations such as SURF. Thus, the paradigm can be admitted as promising.