On the accuracy of rotation invariant wavelet-based moments applied to recognize traditional thai musical instruments

Abstract

Rotation invariant moments constitute an important technique applicable to a versatile number of applications associated with pattern recognition. However, although the moment descriptors are invariant with regard to spatial transformations, in practice the spatial transformation themselves, affect the invariance. This phenomenon jeopardizes the quality of pattern recognition. Therefore, this paper presents an experimental analysis of the accuracy and the efficiency of discrimination under the impact of the rotation. We evaluate experimentally the behavior of the noise induced by the rotation versus the most popular basis functions based on wavelets. As an example, we consider a particular but interesting case of the Thai traditional musical instruments. Finally, We present a semi heuristic pre computing technique to construct a set of descriptors suitable for discrimination under the impact of the spatial transformation.