How to prepare images to better results of accuracy? Make Black/white images? Is Size of image matter?

[SUPERustam]

I have dataset of images and I can make some preprocessing image to reach better results of accuracy. I use Inference model (just use out-of-box PaddleOCR), but I want increase accuracy rate on my dataset. I know I can fine-tune model, but which other methods I can use to reach my goal?