Address Detection in Sanborn Maps with Image Processing and OCR

EDR Mathematics, 2017-18

Liaison(s): Zachary Fisk, Paul Schiffer, Richard White
Advisor(s): Rachel Levy
Students(s): Daniel Zhang, Jeff Carney, Mehdi Drissi, Jordan Haack

Our task is to automatically detect and read the handwritten addresses from EDR’s collection of 1.2 million Sanborn maps. Sanborn maps are detailed city maps produced regularly between 1880 and 2006. We use various image processing techniques to first find the street segments, and then detect the handwritten street names and house numbers. We then run these images through our OCR model, which reliably parses connected characters that are rotated or skewed.