Document Type



Doctor of Philosophy


Computer Science

First Adviser

Lopresti, Daniel

Other advisers/committee members

Huang, Xiaolei; Davison, Brian; Nagy, George; Cao, Huaigu


Many pre-processing techniques that normalize artifacts and clean noise induce anomalies due to discretization of the document image. Important information that could be used at later stages may be lost. A proposed composite-model framework takes into account pre-printed information, user-added data, and digitization characteristics. Its benefits are demonstrated by experiments with statistically significant results. Separating pre-printed ruling lines from user-added handwriting shows how ruling lines impact people's handwriting and how they can be exploited for identifying writers. Ruling line detection based on multi-line linear regression reduces the mean error of counting them from 0.10 to 0.03, 6.70 to 0.06, and 0.13 to 0.02, com- pared to an HMM-based approach on three standard test datasets, thereby reducing human correction time by 50%, 83%, and 72% on average. On 61 page images from 16 rule-form templates, the precision and recall of form cell recognition are increased by 2.7% and 3.7%, compared to a cross-matrix approach. Compensating for and exploiting ruling lines during feature extraction rather than pre-processing raises the writer identification accuracy from 61.2% to 67.7% on a 61-writer noisy Arabic dataset. Similarly, counteracting page-wise skew by subtracting it or transforming contours in a continuous coordinate system during feature extraction improves the writer identification accuracy. An implementation study of contour-hinge features reveals that utilizing the full probabilistic probability distribution function matrix improves the writer identification accuracy from 74.9% to 79.5%.