matthewearl/deep-anpr — reverse-engineered prompt

Build me a simple experimental Python tool that can recognize number plates in photos using a neural network. I want it to be more of a tinkering project than a polished commercial app.

It should let me prepare background images, generate fake UK style number plate training images using a plate font, train the model, save the trained weights, then run detection on a normal input photo and write a new image showing the detected plate. Please make the workflow clear enough that I can run each step from the command line without guessing what to do.

Use TensorFlow, OpenCV, and NumPy if that matches the best approach. Include sensible defaults, clear error messages for missing fonts or image folders, and a README that explains the setup, the big data download, training expectations, and an example detection command.

If any APIs have changed, look up current docs online and adjust the code so it actually runs.