Here is a workflow of automatic reading of a water meter.

An old Android phone is fixed in front of a water meter. "IP Webcam Pro" works as both a camera app and a server with REST interface. A cliant computer periodically runs a bash script (meter.sh) with launchd and fetches camera pictures.

Extract region of interst with imagemagick. A Python+OpenCV script (preprocessor) converts a multi-digit image into a list of single-digit images. Demonstration of this part is explained in OpenCV Water Meter.ipynb.

Manually prepare a CSV file of (image filename, human-recognized reading) pairs. Then a python script (OpenCV Water Meter.ipynb) prepares labels for single-digit images.

Train a neural network using Tensorflow. The procedure is exactly the same as the MNIST recognition in the tutorial.