Skip to content

Latest commit

 

History

History
26 lines (18 loc) · 1.71 KB

README.md

File metadata and controls

26 lines (18 loc) · 1.71 KB

Napoleonic Venetian cadaster processing

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

Segmentation (dhSegment)

The segmentation uses a fully convolutional neural network, dhSegment, which is available at github.com/dhlab-epfl/dhSegment

Transcription (CRNN)

The code for training the convolutional recurrent neural network used to transcribe the labels can be found at github.com/solivr/tf-crnn

Usage

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

Requirements

Create a new environement using the requirements.txt.
Then install dhSegment and tf-crnn.