opencv traincascade positive and negative image

edit

Hello all, I'm using Haar Cascade for detect plate localization. I had using 4500 images for training and 100 other images for testing but result is not good. Probability correct value is only 70%. Currently i have 1500 Positive images and 3500 Negative images. I had using an image contains plate for Positive image and then i remove plate in this image using Paint tool for Negative image. I don't know this way have correct? Does anyone have experience about this?

Positive Image:

Negative Image: