Questions about format of licenseplate

[jeungmin717]

@sirius-ai Hello, Thanks for this wonderful repository

In my case, KOR licenseplate has 7digits or 8digits. Type A , and Type B

Type A is like 12A1234 Type B is like 123B1234 the formats are 2-3 digits + One Letter + 4 digits

However, for some case it would return bad precision

Case1.

target : 36루5854 -> infer: 36루구5854

Case2.

target:97로6695 -> infer: 97로669두

Both cases have wrong format. this cannot be licenseplate format. By this case "my model" has not learn the format as I mentioned earlier, format has only one/two Letters between digits. no letters after digits.

I looked into model and found global context.

Q1. I found code for global context in lprnet model It has already been applied. So just training this model would get advantages of global context (such as formatting). Am I wrong? Should I need some postprocessing functions for formatting?

Q2. Overfitting with this model , Ideas to deal with train data.

I have trained this model with only real car image datas 10k . Got about 90-95% percent Greed_decode_Eval Accuracy But, when I apply this model to general cases, significantly drops to 80% accuracy ( even with good image : clean and nice view of licenseplate)

I think maybe the problem is overfitting , Since it has bad accuracy when tested with other license plate number images

Generated Fake License Plate Image (10k) + Real Images(10k) + Augmented Real Images(30k) training would help for this case? I really need help. Thanks