Offline writer verification based on forensic expertise: Analyzing multiple characters by combining the shape and advanced pen pressure information

Offline writer verification is an important security measure and is related to human traits. However, even with modern techniques, maximizing the amount of information from the limited number of available samples remains a challenging task. We propose a new offline writer verification system inspired by the expertise of forensic document examiners. The system combines the shape and pen pressure information in multiple characters captured by a multiband image scanner. From the visible images, the shape information is extracted as a weighted direction code histogram, and the local pen pressure information (which is based on ink intensity) is extracted as local directional pattern features with adaptive zoning. From the infrared images, the global pen pressure information (which is based on the writing indentations) is extracted as first- and second-order texture statistics. In an experimental comparison study at less than 0.1% error rate, the proposed system reduced the number of required Japanese Kanji characters from six (in the conventional system) to three. After combining these techniques, the error rate per character (averaged from 54 volunteers) was reduced from 4.8% in the conventional method to 1.3%.