We propose the use of a Kohonen Associative Memory (KAM) with Selective Multiresolution (SMR) for the task of modeling and classification of shapes. We demonstrate this technique for Optical Character Recognition (OCR). Modeling is performed on characters selected from 200 fonts as well as documents containing characters with topological shape mutilations, fragmentations (broken characters), and fusions (touching characters). A total of 110,000 training samples are used. This large training set attempts to represent the variation of character shapes due to different font styles, document skew, noise, photometric effects, etc. An omnifont classifier produced using the SMR modeling procedure outperforms a state of the art OCR system. Comparisons to state of the art and benchmark model building procedures are provided.