add noise to generated fake license plates same as real-world taken pics

Question

This is my first time using q&a website to ask. So my question is as follow:

saying I have a license plate image like fig1 generated according how the character layouts, and I've done some resize and gaussian blur jobs with cv2 using python. But the actual real world license plate are like fig2.

Ignoring the layout differences of the characters in the license plate crop, by what kind of operation can I add the noise like fig2?

I have about 4000 image crops like fig2.

Attempts have been made.

• I tried blur method with all kinds of filter and didn't work;
• I tried to blur the MJSynth dataset and trained a CRNN network to recognize my real license plates, didn't work either. I haven't tried directly using real license plates to train since I haven't labeled them and think there might be a better solution.

and I've searched a little about GAN, thinking that maybe GAN or similar networks can help me learn the noise pattern of the real license plate, and afterwards add it to the fake ones to train. but just don't know where to start.

My point is, if I can generate photos like fig2 with label, then I may directly feed them into a CRNN network that's been written to recognize the numbers inside the license plate crop. So again, how can I add noise like fig2 above? both deep-learning instructions and simple morphological steps or any other codes are welcome.

Thanks all!