How can I improve a tensorflow model to recognize vehicle registration plate detection?

I have struggled for a week to train a model that can give me the bounding numbers for extraction of vehicle registration plates. (It's a law enforcement initiative.) The latest attempt - there were many - went like this:

1. I used the pets pre-trained model.
2. I took images of roads and overlaid a different registration plate on each road image. The plate is always in the same position to simplify the generation of the xml files.
3. Upsized the plates to 300x80 - previously much smaller with similar results.
4. 1600 training images and 200 testing images.
5. Converted the xml to csv and then to tfrecord format.
6. Trained the model using the ssd_mobilenet_v1_pets.config with minor changes, such as the num_classes and num_examples and paths.
7. Trained on an AWS GPU instance for a few hours.

Results are horrible. Windscreens are detected as plates and square fronts of vehicles are detected. I can upload if it will help.

I'm about to give up and use a text extraction algorithm, but that brings other complexities such as ignoring other text in the image.

I hate giving up, so any advice eagerly accepted.

Thanks, Alan.

Ouch. Down-voted already. Apologies. Marked as answered. Guess I misunderstood how to formulate questions. Plan B it is.