Orthonormal wavelet representations for recognizing complex annotations

Abstract

This paper describes a novel method of pattern recognition targeted for recognizing complex annotations found in paper documents. Our investigation is motivated by the high reliability required for accomplishing autonomous interpretation of maps and engineering drawings. The recognition problem is made difficult in part because characters and text may be expressed in arbitrary fonts and orientations. Our approach includes a novel incremental strategy based on the multiscale representation of wavelet decompositions. Our approach is motivated by biological mechanisms of the human visual system. Choosing wavelets that are simultaneously localized in both space and frequency, and decomposing a signal into a multiscale hierarchical basis with orientation selectivity, can provide a powerful methodology for pattern analysis. We evaluated several wavelets with different spatial-frequency characteristics and measured their performance in the context of character recognition. Wavelet bases are more attractive than traditional hierarchical bases because they are orthonormal, linear, continuous, and continuously invertible. The multiscale representation of wavelet transforms provides a mathematically coherent basis for multigrid techniques. In contrast to previous adhoc approaches, our method promises a practical solution embedded in a unified mathematical theory. A feasibility study is described in which more than 10000 patterns were recognized with an error rate of 2.6% by a neural network trained using multiscale representations from a class of 52 distinct alphanumeric patterns and graphical symbols. We observed a 10-fold reduction in the amount of information needed to represent each character for recognition. These results suggest that high reliability is possible at a reduced cost of representation.