On Rotational Invariance for Texture Recognition
Xiaojin Shi, Roberto Manduchi · 2005
In this paper we analyze the effect of rotational invariant operators for texture recognition using cross-validation experiments with different sample sizes. This work presents three main contributions. First, invariant operators for steerable filter banks are derived analytically using the Lie group approach. Second, the use of “randomized invariants” for steerable texture analysis is introduced. Randomized invariants produce classification rates that are intermediate between those of non-invariant and of invariant features. Third, a thorough quantitative analysis is presented, highlighting the relationship between classification performances and training sample size, textural characteristics of the data set, and classification algorithm. 1.