In short: Is it needed or not to improve accuracy?
A bit longer: I was going through documentation and Internet and I did not find any references concerning Mobile Vision and if it is doing some sort of image pre-processing procedures by itself or not. I am talking about e.g. deskewing, binarisation or etc.
I know that e.g. you might improve tesseract-ocr accuracy by pre-processing images.
The TextRecogniser does pretty well without any prosessering, but from my (very limited) experiments it does not seem to do any deskewing. If you rotate the phone by >30 degrees the recognition breaks down and if have artifacts very close to the text then it also seems to break down.
Tested it with the lab from: https://codelabs.developers.google.com/codelabs/mobile-vision-ocr/