for the structure and table problem, vision LLMs are worth considerng alongside traditional OCR. paddleocr is a solid step up for raw text accuracy, but for tables, form layouts, and messy formatting it still needs post-processing. a multmodal model handles cleanup and structuring in the same call.

rough cost with haiku: around $0.002-0.005 per page dependng on complexity, which competes well with google vision for low to medium volume:

import anthropic, base64, io
from pdf2image import convert_from_path

def ocr_page_with_claude(pdf_path, page_num=0):
    images = convert_from_path(pdf_path, first_page=page_num+1, last_page=page_num+1)
    buf = io.BytesIO()
    images[0].save(buf, format="PNG")
    img_b64 = base64.standard_b64encode(buf.getvalue()).decode()

    client = anthropic.Anthropic()
    msg = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=2048,
        messages=[{"role": "user", "content": [
            {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": img_b64}},
            {"type": "text", "text": "Extract all text. Preserve tables as markdown. Return only the extracted content."}
        ]}]
    )
    return msg.content[0].text

main downside: latency is higher than local OCR so batch if you have volume. but for the "output needs cleanup" problem it solves two steps at once

paddleOCR is a big step up from tesseract for scan quality, especially anything with slight rotation or uneven lighting.

but the preprocessing before the OCR maters just as much as which engine you pick:

import cv2
import numpy as np
from pdf2image import convert\_from\_path

def preprocess\_page(img):
    gray = cv2.cvtColor(img, cv2.COLOR\_RGB2GRAY)
    denoised = cv2.fastNlMeansDenoising(gray, h=10)
    \_, binary = cv2.threshold(denoised, 0, 255, cv2.THRESH\_BINARY + cv2.THRESH\_OTSU)
    return binary

pages = convert\_from\_path("exam.pdf", dpi=300)

dpi=300 is the big one. most people set it to 150 and wonder why small text comes out garbled. under 200 both engines strugle.

for tables in exam papers: paddle's table detection works ok but you'll probaly still end up postprocessing the bounding boxes for anything structural. docling handles layout beter if tables become the bottleneck.

google vision skip for now. paddle + decent preprocessing gets you to 85-90% on typical exam scans.