A PDF's text layer and its scanned images

A PDF's text layer covers its vector text. The scanned table pasted into the middle of the page is a bitmap, and getTextContent cannot see a word of it.

The usual advice — decide per page whether to trust the text layer or OCR the whole raster — gets the hybrid page wrong both ways. The text layer misses the scan; whole-page OCR re-reads clean vector text worse than the PDF already states it.

Read both

import { Pipeline } from '@stabrise/scaledp'
import { PdfToDocument, PdfEmbeddedImages, PdfMergeImageText } from '@stabrise/scaledp/pdf'
import { PaddleTextRecognizer } from '@stabrise/scaledp/ocr'

const pipeline = new Pipeline([
    new PdfToDocument({ resolution: 300 }),
    new PdfEmbeddedImages({ resolution: 300 }),
    new PaddleTextRecognizer({ inputCol: 'image', outputCol: 'image_text' }),
    new PdfMergeImageText(),
])

const pages = await pipeline.transform(file)

One row per page comes out, and document.bboxes holds the text-layer words and the words read out of the page's images together, in one coordinate space, in reading order.

What each stage does

1. PdfToDocument reads the text layer into word boxes and sets the page index on the row.
2. PdfEmbeddedImages extracts that page's raster images at their own pixel size — not a crop of the rendered page, so nothing is resampled — and emits a row each, with a placement saying where the image sits.
3. Any recognizer reads them. They are ordinary images; nothing about this step is PDF-specific.
4. PdfMergeImageText maps each image's boxes back onto the page through its placement, merges them with the text layer, and reduces the page's rows back to one.

The row count goes 1 → pages → images → pages. Every PDF stage honours a page index the row already carries, so the two readers do not multiply.

Overlapping readings

A searchable scan — a photograph with an invisible OCR layer over it — describes the same words twice. Two things handle it:

• PdfEmbeddedImages skips an image that already has minCoveringBoxes text-layer boxes inside it, so the wasted recognizer pass never happens.
• PdfMergeImageText's strategy decides the rest. text-layer-wins is the default; ocr-wins is for corpora whose text layer extracts as mojibake.

When you only want the cheap answer

If your corpus really is either-or, the two-pass split is still cheaper — see Skip OCR when the PDF already has text.