Some datasets can be used to train Automatic License Plate Recognition (ALPR).
- Caltech Cars
- English License Plate
- UCSD-Stills (UCSD car dataset)
- CCPD: Chinese City Parking Dataset
- Holistic Recognition of Low Quality License Plates by CNN using Track Annotated Data
- OpenALPR benchmark
- 500 Images of the Rear View
A quick glances at those datasets is :
In a short word, the following datasets are large enough to be used in building ALPR
There’re other sites that can we use to mined license plate images such as:
The following open-source softwares are available to use in commercial ALPR:
- ultimateALPR-SDK : Run very fast even for embedded devices(105fps on GTX 1070, 47fps on Snapdragon 855, 12fps on Raspberry Pi 4) . The source code can run locally. The price is at pricing. They also has the web demo at webapp
- openalpr : The most famous ALPR software. It was built on opencv and tesseract, so when comparing to deep learning approach it is not that strong. The license is AGPL-3.0 so it’s quite nuissance when we want to use it in commercial.
A quick search in github can give some project that use deep learning for ALPR. Unfortunately, those projects are all half-done and not suitable for product.
There’re quick report about openalpr’s performance inside the paper as follow:
The ultimateALPR-SDK project is amazing. It’s worth a try.
3. Personal thinking
The performance of ultimateALPR-SDK is amazing. The price is a little tricky. If we develop a webapp then only the server license is needed and it is very cheap. But if we want to develop on edge then we need a developer license which price is hidden. Maybe in the future, when i have chance, i will contact with them to get more detail about developer license.