Find signatures and faces

SignatureDetector and FaceDetector are YoloOnnxDetector with a model and labels already chosen. They read a page image and write a box column.

import { Pipeline, ImageDrawBoxes } from '@stabrise/scaledp'
import { PdfToImage } from '@stabrise/scaledp/pdf'
import { FaceDetector, SignatureDetector } from '@stabrise/scaledp/detect'

const pipeline = new Pipeline([
    new PdfToImage({ resolution: 200 }),
    new SignatureDetector({ scoreThreshold: 0.25 }),
    new FaceDetector(),
    new ImageDrawBoxes({ inputCols: ['image', 'signatures'], outputCol: 'pass1', color: '#9d8cff' }),
    new ImageDrawBoxes({ inputCols: ['pass1', 'faces'], outputCol: 'annotated', color: '#ff5c8a' }),
])

They write different columns by default — signatures and faces — so both can run in one pipeline without colliding.

200 DPI is enough here. These are object detectors on page-scale features, not text recognition, and rendering at 300 costs time for no gain.

Cropping each one out

import { ImageCropBoxes } from '@stabrise/scaledp'

new Pipeline([
    new PdfToImage({ resolution: 200 }),
    new SignatureDetector(),
    new ImageCropBoxes({ inputCols: ['image', 'signatures'], padding: 8, returnEmpty: false }),
])

One row per signature, with the crop in cropped_image and its source box in box. padding: 8 gives the crop a margin — a tight YOLO box usually clips the tail of a signature.

returnEmpty: false (the default) records No boxes to crop in the row's exception when a page has none, rather than emitting the whole page. Whether that is a failure or a result depends on your job; set it to true if "no signature on this page" is expected.

Blacking out faces

new ImageDrawBoxes({
    inputCols: ['image', 'faces'],
    outputCol: 'redacted',
    filled: true,
    color: '#000000',
    padding: 4,
})

Same shape as PII redaction, and the same caveat: it paints the rendered image, not the source PDF.

Tuning

scoreThreshold defaults to 0.2 on both, which is deliberately permissive — these detectors are usually used to route a document to a human, where a false positive is cheap and a miss is not. Raise it for automated filing.

iouThreshold (0.5) suppresses duplicate detections per class. Lower it if a single large signature comes back as two overlapping boxes.

Another YOLO model

Anything you have as a YOLO ONNX export runs through the base class:

import { YoloOnnxDetector } from '@stabrise/scaledp/detect'

new YoloOnnxDetector({
    model: 'my-org/stamp-detection',
    labels: ['stamp', 'seal'],
    outputCol: 'stamps',
    outputType: 'stamp',
})

Both the transposed YOLOv8/v11 layout and a graph with NMS baked in are decoded. See YoloOnnxDetector.