This work was presented at the Digital Humanities 2019 conference in Utrecht, A deep learning approach to Cadastral Computing.
To see the old version of the work presented in 2017 (Machine vision algorithms on cadaster plans), go to this branch
The segmentation uses a fully convolutional neural network, dhSegment, which is available at github.com/dhlab-epfl/dhSegment
The code for training the convolutional recurrent neural network used to transcribe the labels can be found at github.com/solivr/tf-crnn
In order to use this code, two models need to be trained: a segmentation model (dhSegment) and a transcription model (CRNN). Once the models are trained, it is possible to process the casdaster maps by running:
python process_cadaster.py -im image1.tif image2.jpg -out output_dir -sm segmentation_model_path -tm transcription_model_path
-im: image(s) to process (can be of type .tif, .jpg or .png)
-out: output directory where the geojson containing the geometries and their transcriptions will be saved
-sm: path to the Tensorflow saved model for segmentation
-tm: path to the Tensorflow saved model for transcrition
Create a new environement using the requirements.txt.
Then install dhSegment and tf-crnn.