Automated signature verification is an important capability when doing fraud detection from high volumes of document images. The challenge is mainly caused by the combination of inherent writer variations and the conscious attention placed by the fraudster on imitating a legitimate signature. Getting good signature references is the first and most important step affecting verification accuracy. Typical production systems use signature cards which have one or more signature variations. In this paper, we show the improvement in verification accuracy from increasing the number of signature references used in verification processing. As many as 20 signature references were used. In production operations, comparing a sample signature to more than one or two reference signatures using current methods requires too much processing time. The main contribution of this paper is the description of a feature-based approach that allows for matching of a sample signature against a large number of reference signatures in real-time thereby achieving improved accuracy without a significant performance penalty.