LineOrientationDetector

Import

import { LineOrientationDetector } from '@stabrise/scaledp/ocr'

Group

Transform

Reads

image, boxes

Writes

image, orientations

Needs

onnxruntime-web

Note

Expands one row into several.

A port of ScaleDP's HasDetectLineOrientation, which there is a mixin on TesseractRecognizer. Here it is a stage, so it can sit in front of any recognizer.

import { Pipeline } from '@stabrise/scaledp'
import { PdfToImage } from '@stabrise/scaledp/pdf'
import { DbnetOnnxDetector, LineOrientationDetector, PaddleTextRecognizer } from '@stabrise/scaledp/ocr'

const rows = await new Pipeline([
    new PdfToImage(),
    new DbnetOnnxDetector({ outputCol: 'boxes' }),
    new LineOrientationDetector({ inputCols: ['image', 'boxes'] }),
    new PaddleTextRecognizer({ inputCol: 'oriented' }),
]).transform(file)

It needs boxes, so it goes between a detector and a recognizer. It writes two columns: the corrected image to outputCol, and one '0_degree' / '180_degree' label per box to orientationCol.

If nothing was flipped, the original image object is passed through without being re-encoded.

Why a stage rather than a mixin

Python corrects each crop inside the recognizer. PaddleTextRecognizer detects and recognises in one pass and offers no seam to hook into, so correction has to happen before it.

Turning each inverted region in place on a copy of the page reaches the same result with one recognition pass rather than one per box. It works because a rectangle maps onto itself under a 180° rotation about its own centre, so every box's coordinates stay valid afterwards. The turn is clipped to the box's own rotated outline; without that, a skewed box drags its neighbours through the rotation with it.

onlyRotated, measured rather than assumed

The cost of that default is the case it misses: an upside-down horizontal line. Verified on a page with one, onlyRotated: false recovers it — Th sd s ie own becomes This line is upside down — while the default skips it.

Turn the gate off when a document may contain inverted horizontal text, and leave it on otherwise.

correct: false classifies without touching the image, which is the cheap way to find out whether a corpus needs correcting at all.

Parameters