Neural network pattern recognition employing multicriteria extracted from signal projections in multiple transform domains

We propose a novel one and multidimensional signal classification system that employs a set of criteria extracted from the signal representation in different transform domains, denoted the multicriteria multitransform (MCMT) classifier. The signal projection, in each appropriately selected transform domain, reveals unique signal characteristics. These characteristics in the different domains are properly formulated to obtain classification criteria with efficient implementation properties such as speed and accuracy. Results for image classification confirm the improved classification performance relative to existing techniques. In addition to the improved computational efficiency, the proposed technique maintains higher classification accuracy in the presence of additive noise.