Adding Feedback to Improve Segmentation and Recognition of

WinBank is a system which performs automated reading of handwritten documents, particularly of strings of handwritten numerals on bank checks. A commonly-used strategy for reading handwritten strings is a combination of segmentation and recognition. In the WinBank program, segmentation involves both the separation of touching characters and the merging of character fragments to other pieces. An inherent problem with the segmentation/recognition strategy is the fact that if segmentation is done incorrectly, the erroneous characters cannot be recognized. This paper discusses a new method for feeding rejected characters back into the segmentation process to perform error recovery. Such feedback reduces the frequency with which segmentation errors occur, and increases the probability that each character will be recognized. Thesis Supervisor: Dr. Amar Gupta Title: Co-Director, Productivity From Information Technology (PROFIT) Initiative Sloan School of Management