madelonhulsebos / ComputerVision

This repository comprises the work for an academic project in the context of Computer Vision. We built a system that reads in a file, scans an image in search for a license plate, and if one was found outputs the characters on this license plate.
6 stars 5 forks source link

Construct proper training datasets for license plate detection #4

Closed JeroenMandersloot closed 7 years ago

JeroenMandersloot commented 7 years ago

At http://www.acme.com/licensemaker/ we can generate synthetical license plates. I have already written a script that downloads randomly generated license plates for various states from this site, but we may still need to do some additional work to make them usable as training samples. For example, we may need to add background noise before we can train a classifier which can identify whether or not a license plate is present in a sliding window extract.

JeroenMandersloot commented 7 years ago

To add backgrounds to the synthetic license plates, I have downloaded the SUN image dataset and selected random images to be used as a background. Right now I have generated about 2,500 images using this procedure, and it's pretty easy to create even more. Below I have depicted some of the typical results of my current procedure:

9mm6vmb_delaware 0vaosn1_oklahoma 4v7c0l5_illinois

I have yet to test if these work well as training samples in a practical setting. If not, there are still several other things I could try to refine this technique: