An investigation on the use of local multi-resolution patterns for image classification

• This paper investigates the use of local multi-dimensional patterns for image classification.
• We present a rigorous and general mathematical full model for encoding multi-resolution, rotation-invariant local patterns.
• We evaluate the use of multi-resolution patterns for image classification both from an information- and performance-based standpoint.

The subject of this study is the use of local multi-dimensional patterns for image classification. The contribution is both theoretical and experimental: on the one hand the paper introduces a complete and general mathematical model for encoding multi-resolution, rotation-invariant local patterns; on the other experimentally evaluates the use of multi-resolution patterns for image classification both from an information- and performance-based standpoint. The results indicate that the joint multi-resolution model proposed in the paper can actually convey an additional amount of information with respect to the marginal model; but also that the marginal model (i.e. concatenation of features computed at different resolutions) can be a good enough approximation for practical applications.