Can we use Vision API + NLP to read text from image and categorize like Cloud Vision API ?

Nope. Vision only gives you the rectangles that contain text, it does not have an API to convert these image regions to text.

Actually it's possible by adding a CoreML model such as MNIST.

So, with Vision, you detect the bounding box, then you extract the image portion inside the bounding box and you give it to MNIST model.

One precompiled model for CoreML is here:

https://github.com/ph1ps/MNIST-CoreML

Take a look at my blog:

neurosurg dot de

It also includes training and samples ..