This dataset was created by first training an object detection to detect license plates. They were all then cropped based on the bounding box output.

The original license plates were taken from a synthetic dataset which can be found here https://www.kaggle.com/datasets/tbjorklund/annotated-synthetic-license-plates and the credit for that goes to TBjorklund.

Unfortunately the dataset is missing instances of two letters but it is still great for experimentation. There are about a handful of images that were cropped incorrectly, <1%.