Detect, then recognise

PaddleTextRecognizer detects and reads in a single pass, which is fast and convenient and gives you no say in the geometry. When the page is skewed, rotated or unusually laid out, split the two jobs.

import { Pipeline, ImageDrawBoxes } from '@stabrise/scaledp'
import { PdfToImage } from '@stabrise/scaledp/pdf'
import { DbnetOnnxDetector, TesseractRecognizer } from '@stabrise/scaledp/ocr'

const pipeline = new Pipeline([
    new PdfToImage({ resolution: 300 }),
    new DbnetOnnxDetector({ outputCol: 'boxes', unclipRatio: 2.5 }),
    new TesseractRecognizer({
        inputCols: ['image', 'boxes'],
        boxLevel: 'word',
        padding: 5,
    }),
    new ImageDrawBoxes({ inputCols: ['image', 'text'], outputCol: 'annotated', lineWidth: 2 }),
])

Why TesseractRecognizer and not the Paddle one

Because it is the only recognizer that reads boxes it is given. PaddleTextRecognizer ignores an upstream box column entirely — the boxes it returns are the ones its own detector found. Putting a detector in front of it gives you two opinions about the page, not a pipeline.

Add orientation correction

If lines may be upside down, put the classifier between the two:

new Pipeline([
    new PdfToImage(),
    new DbnetOnnxDetector({ outputCol: 'boxes' }),
    new LineOrientationDetector({ inputCols: ['image', 'boxes'] }),
    new TesseractRecognizer({ inputCols: ['oriented', 'boxes'] }),
])

The recognizer now reads the corrected page. Box coordinates survive the correction because a rectangle maps onto itself under a 180° turn about its own centre.

TesseractRecognizer can also do this itself — detectLineOrientation is on by default and classifies each crop. The separate stage is what you want when the recognizer is Paddle's, which offers no such seam.

Word boxes out of line-level detectors

Every detector here is line-level, in Python ScaleDP as here. boxLevel: 'word' maps Tesseract's own word rects back through the crop — padding, scaleFactor and rotation included — so a word inside a skewed line comes back skewed the same way.

That is what makes entity redaction precise rather than line-shaped.

Comparing detectors

Run two and draw them in different colours:

new Pipeline([
    new PdfToImage(),
    new PaddleTextDetector({ outputCol: 'paddle_boxes' }),
    new DbnetOnnxDetector({ outputCol: 'dbnet_boxes' }),
    new ImageDrawBoxes({ inputCols: ['image', 'paddle_boxes'], outputCol: 'pass1', color: '#3fc9f5' }),
    new ImageDrawBoxes({ inputCols: ['pass1', 'dbnet_boxes'], outputCol: 'annotated', color: '#ff5c8a' }),
])

Two passes rather than one, because a single ImageDrawBoxes takes one colour for all its sources.