ANPR Data Sets For Deep Learning

Question

I am working on an ANPR application for UAE license plates which differ themselves from one another in shapes and partial different colours. As to make a prototype I like to use transfer learning with around 1000 images of license plates. My first step is to detect license plate and then extract it.

For that I have some doubts:

1. How should the image of license plate be fed for training? Should it contain license plate only or license plate with some other parts of the car?
2. What should be the minimum size of an image regarding bytes? Is it okay if I use an image of few KiloBytes?

Answer 1

It all depends on your application/learning framework. For example some frameworks will do the resize for you.

How should the image of license plate be fed for training? Should it contain license plate only or license plate with some other parts of the car?

Anyway, based on my experience with licence plate recognition for Arabic letters I would suggest to include a thin outer border of plates for more accuracy. There will be good features in that lines.

Example:

What should be the minimum size of an image regarding bytes? Is it okay if I use an image of few KiloBytes?

Clear images around 52 x 150 pixels will do. Again, it depends on your implementation, but generally speaking larger images made no difference in precision, but will increase over-fitting risk and training time.