If anyone is up for creating a car number plate detector based on machine learning (such as neural networks), tons of images for training is required. There is good amount of data available online for training number plate detectors. However, here's a problem - they do not work very well with Bangladeshi number plates; it's because Bangladeshi number plates are quite different from British, American, or even Indian ones. Images of Bangladeshi number plates are extremely limited on Google, and no dataset containing such images is publicly available. This calls for the need of creating a dataset containing photos of Bangladeshi number plates, which will be released publicly for free. This is intended to help everyone from beginners to machine learning to data scientists. Our work is intended to be primarily used for training neural networks.
To create a good dataset, the data needs to be diverse and distinguishable. Too similar data can be misleading to a neural network. Here are few instructions on helping you capture images which will make it effective for a neural network to learn from:
- Make sure the image contains number plate of the car clearly identifiable; it's okay if the number plate is partially visible as it'll give challenges to the neural network, but make sure you do not take too many photos of partial number plates.
- Take photos with number plate taking up 10% of the image in few shots, 20% in the next few ones, and so on. Try to keep up the diversity.
- Take photos of the number plate from various angles; e.g. one from front, diagonally left, diagonally right, rear, and so on.
- Take photos at varying lighting setups; e.g. take at morning, evening, dusk, night, etc. Do not forget that lighting setup varies with location; for example, lighting at a basement is likely to be totally different from lighting in open road.
- Do not worry if some of the images are blurry, but make sure that the percentage doesn't get too high; proper balance is the key here.
- Try not to capture similar images; for example, it's likely that you'll stand facing an open road, tapping your phone to capture photos; just make sure you do not capture photos of the same number plate, in the same angle multiple times; it only contributes to redundancy of the dataset.
Rear view:
