Extraction of Handwriting in Tabular Document Images

We propose a method for detecting handwriting in sets of tabular document images that share a common form. This is accomplished leveraging previous work on aligning structured documents. These aligned documents are processed together as an image stack. First, the blank form common to the documents is generated using image averaging. Then, each image is compared individually to the blank form, and regions with a large difference are marked as handwriting. Results are pursued under the assumption that a good handwriting detection algorithm will have as few false positives as possible while maximizing recall. Proof of concept efforts are convincing, though a more indepth analysis remains to be done. Work to filter out false positives will be pursued.