v3.1.1

What's new in v3.1.1

findRegions() — thresh option for resized binary images

The default thresh: 127 caused accuracy loss when running findRegions on a resized binary image. Resizing introduces anti-aliased border pixels (values 1–127) that the old threshold ignored, while OpenCV's findContours treats any non-zero pixel as foreground.

Use thresh: 0 to match OpenCV behaviour:

const regions = new CanvasProcessor(resizedBinaryCanvas).findRegions({
  foreground: "light",
  thresh: 0,           // any non-zero pixel = foreground (matches OpenCV)
  minArea: 20,
  padding: { vertical: 0.4, horizontal: 0.6 },
  scale: 1 / resizeRatio,
});

With thresh: 0 + padding + scale, full-pipeline IoU vs OpenCV is 98.4% (all 21/21 boxes matched).

New findRegions() options

Option	Default	Description
thresh	127	Foreground threshold. Use 0 to treat any non-zero pixel as foreground (matches OpenCV on resized binary images).
padding	—	{ vertical?, horizontal? } — expand each bbox by a fraction of its height, mirroring applyPaddingToRect.
scale	1	Multiply all coordinates by this factor (use 1 / resizeRatio to map from processed→original image space).

New example

examples/find-region-vs-get-contours.ts — side-by-side visual comparison of CanvasProcessor.findRegions vs OpenCV contours on a real receipt image, producing annotated PNG output files.

Type safety

All comparison.test.ts type errors resolved without as any. Canvas-to-canvas copies now route through CanvasProcessor.resize to avoid @napi-rs/canvas drawImage typing constraints.