This paper describes a technique for classification of 2-D discrete signals. It consists of four modules, namely the partition, representation, measurement, and the classification modules. The first of these either passes the observed signal as a whole or divides it into subregions which may or may not overlap. The representation module first computes the shift-invariant multiscale wavelet representations (MSWAR) of the reference and the observed signals and then generates a corresponding set of 1-D signatures. The measurement module extracts those vital signal features to which the decision rules of the classification module are applied. The paper presents the design and implementation of each of these modules, emphasizing theoretical background behind the design and efficiency of their implementation. Also some preliminary results have been included that demonstrate the ability of this technique to classify observed signals that are corrupted by different types of deformities.