Digitize Scanned Contour Map with deep learning

[pyGisServer]

Hello to all the community, I hope you are well!

I have been trying for a long time to find a method with which to digitize contour maps like the one in the following image:Contour Map Tile Sample

. So far the best I've found is to use ArcScan automatic vectorization ,RasterToPolyline (Conversion)  , or third-party software with plenty of post-processing and in most cases obtaining unsatisfactory results. With Deep learning using the ArcGIS Image Analyst extension I started to investigate the new features and consider it a perfect task to test its limits, since I think it would be the ideal approach to achieve a model that is capable of obtaining good results despite the fact that there are disjointed contours, punctually covered by elevation numbers or areas where there are unwanted groupings of pixels for a outline, however, I still don't know how to start training my model. I found something similar in Geo-Referencing and Digitization of Scanned Maps , but despite it looking promising I think it's not meant to extract contours (point me if I'm wrong). I have a large dataset for training, but I don't know how to use it with ArcGis Learn, my dataset has numerous maps in image format and a polyline shapefile with the contours extracted from each of the maps, however, according to the documentation you reviewed, the sample sets should have polygon shapefiles and not polyline shapefiles. However, as far as he was able to investigate, it is possible to train a network to obtain polylines as seen in Automatic road extraction using deep learning 

I would like to know if anyone has tried to do something similar and which Arcgis Learn module you recommend me to address the problem with.

[PavanYadav]

@pyGisServer 

What kind of rasters/images do you have in mind to use for training the model? If it's going to be DEMs, I believe the Contour tool in ArcGIS will be very efficient. From your description it does not look like you will be using DEMs. If so, please ignore this and continue reading. 

But if you need to extract (digitize) lines on topographic map images like you shown in your above example, then yes you can try building a deep learning model. I believe these model architectures might be good for this task: HED Edge Detector, BDCN Edge Detector, and Multi Task Road Extractor. For these model architectures you will use the Classify Pixels for Deep Learning tool and it will output rasters. Finally, like you said in your post, you can convert the rasters to polylines.