How can i train it for two line license plate?

[arsalan993]

Also what about license plate which are out of sample , i mean that they have different font type and size and color and dimension of license plate has little bit variation. and there is no specific pattern inside license plate.

[qinhuan]

My train dataset is china's license plate. Aspect ratio of license plate is about 1:3. Most china's license plate is blue. So this model is more suitable for china's license plate. Cascade CNN was originally used to face detection, aspect ratio of face is about 1:1. Maybe this method is not universal for license plate. It can not handle situations where the aspect ratio varies widely.