S. Benoit, F.P. Ferrie The rich archives of topographical maps, useful for both historical and contemporary reference, are difficult and tedious to transfer to computer representation. Once the contours are digitized and labelled with the correct elevation, a computer can reconstruct the 3D terrain, for purposes of fusion with satellite data or comparison with other sources of terrain information. The images are typically noisy and cluttered with symbols that distract from the contours of interest, which are themselves in poor contrast. This task calls for a robust method of extracting curves from a noisy image, with the ability to filter out unwanted artifacts and symbols. The approach chosen follows from the work of Zucker et al., using a computational model that resembles the hypothesized organization of primate vision. Results obtained by previous work have shown success with similar images. Edge information is extracted using Logical-Linear operators. Edge elements are filtered and refined using Relaxation Labelling (work in progress) using a hypothesis of curvature consistency, the tendency that neighboring elements along a curve support the same curvature. At present, several strategies are being studied for the integration of the curve elements into larger curves. The goal of this project is to develop a system whereby an image of a map is processed, generating a set of elevation curves. An operator would then need to select an elevation for each curve. The final stage of 3D terrain reconstruction is considered a forward projection problem, and many techniques already exist to interpolate a surface given the sparse contour curves. This is not the main concern of the project; the extraction of the elevation contours is the most demanding and significant contribution of this work.