Segmentation of Handwriting on Maps

Given an image of an annotated map, we detect and sement hand-drawn annotations as a pigmented layer, both reflective and translucent. Unlike many methods that merely explore different color spaces, we actually employ a more accurate Kubleka-Munk color mixing model for annotations. In the process, our method analyzies each pixel based on color and texture and represents it as a sequence of layers that can be overlaid to reporduce the original map, with some layers highlighting curvlinear networks that are good candidates for annotation symbols. We explore a new Markov Random Field for lableing pixels in a manner that encourages curvilinear results. Higher level analysis is used to prune pixels from curves that are not annotations. furthermore, we provide tools to generate and process such documetns, for example by applying symbol-spotting techniques to allow particular glyphs to be identified.