PdfEmbeddedImages

Import

import { PdfEmbeddedImages } from '@stabrise/scaledp/pdf'

Group

Read

Reads

bytes

Writes

image, placement

Needs

pdfjs-dist

Note

Expands one row into several.

A PDF's text layer covers its vector text and nothing else. Anything scanned, photographed or pasted in as a bitmap — an invoice table, an ID page, a receipt — is invisible to PdfToDocument. This stage finds those bitmaps and emits one row per image, so any OCR stage can read them like any other picture.

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

const rows = await new Pipeline([
    new PdfToDocument({ resolution: 300 }),
    new PdfEmbeddedImages({ resolution: 300 }),
    new PaddleTextRecognizer({ inputCol: 'image', outputCol: 'image_text' }),
    new PdfMergeImageText(),
]).transform(file)

See A PDF's text layer and its scanned images for the whole pipeline.

The pixels are the PDF's own

The alternative — render the page and crop the image's rectangle out of it — resamples twice and is capped by the render DPI. This stage takes the image object itself, so the pixels reach the recognizer exactly as the PDF stored them.

That is not the same as more pixels, and the distinction matters. In SampleWithFaceImage.pdf the embedded scan is 1524px wide and is placed across 1981px of a page rendered at 300 DPI — an effective 231 DPI. Rendering the page at 600 would not recover detail the scan never had; it would only interpolate.

placement.effectiveResolution reports that number, and image.resolution carries it onto the emitted image. minEffectiveResolution upscales anything below it before the recognizer sees it, which is worth doing: most recognizers are trained around 300 DPI and read a 90 DPI thumbnail badly.

It reads one page, not the whole file again

Two expanding PDF stages over one file used to square the row count — a five-page document through PdfToDocument and then another reader became twenty-five rows. Every PDF stage now honours a page index the row already carries, so a chain reads each page once. Only a PDF stage ever writes that column.

What it skips, and why

• Images smaller than minPixels. A bullet, a rule and a gradient swatch are all images, and each one would otherwise cost a recognizer pass.
• Stencil masks. A 1-bit /ImageMask tints a shape; it carries no readable pixels.
• Images already covered by the text layer. A searchable scan — a photograph with an invisible OCR layer laid over it — has hundreds of text-layer boxes inside the image's placement. Re-reading it costs seconds to arrive at a worse answer than the layer already holds. minCoveringBoxes is that test; set it to 0 to read every image regardless.

A page with no images

returnEmpty (on by default) emits the page carrying an Image whose exception says there was nothing to read. A downstream recognizer reports that faithfully and PdfMergeImageText recognises it as the expected case, so the page keeps its text layer instead of disappearing from the results.

Parameters