Could you give me some suggestions for training dataset?

[yangxiuwu]

Hi, I have trained a Chinese OCR model by CRNN ( 300W synth text image as train dataset). but the model has poor result for the real scene. So could you give me some suggestions for training dataset: Does the dataset require a fixed aspect ratio? Does the dataset need some data augment, e.g. transform , blur, different font color and diverse background and so on?

[Cocoalate]

Hi, is your 300W synth text image data public? I'm working on receipt ocr now but my data hasn't been enough.