I want to train a super-resolution model that I will further apply to text images. The goal is to work with camera-captured images of text, increase the resolution so that the quality of OCR increases. Ideally, the model should both increase the resolution (and get better results compared to bicubic interpolation) and also improve the image quality.
Even though the best way to collect data would be to use a lot of different cameras, various light conditions, etc., I am wondering if it's possible to artificially add camera-specific noise to images. I already have a dataset of mobile camera-captured images of text, but the quality is decent, so I think it makes sense to apply some filters and, well, "make some noise".
From what I have read so far, I could use Gaussian, uniform and salt&pepper in random combinations, followed by median filtering.
What else could I do to at least partially imitate the quality of mobile camera defects?