Data Extraction OCR

I just need model with source code explanation (not looking for end to end project or UI)

I have unstructured pdfs with scanned data, I want convert it into image and improve image quality.
- text skewed in angle in 3d (rotation , curved etc)
- images with different light conditions, and quality
- images with different sizes
Perform ocr and extract fields like
Vendor/Company Name
Address (Shipping/Billing)
TAX
SubTotal
Total

I have very few samples (dataset)
No third party (paid) api should be called, it should be built purely in tensorflow..