Three numbers on the hero are not decoration
Five operators cover the gradients most coursework still assigns by name. Zero uploads means the canvas API reads bytes locally, the same way noise reduction keeps your grain experiments private. PNG export preserves the mask you see, which matters when you paste into a slide or trace in a vector app.
Stat panels also set expectations. This page is a teaching bench, not a replacement for industrial pipelines. You get fast feedback, not sub-pixel hysteresis tuning across octaves.
When a silhouette is the whole story
Packaging teams sometimes need to check whether a logo survives once color is stripped. A quick Sobel pass shows if interior strokes collapse into mush.
Students comparing kernels for a report want side-by-side truth without installing toolchains. Swapping Prewitt for Roberts here shows how diagonal emphasis shifts without rewriting homework scripts.
Photo editors chain steps: if halos appear after sharpen, edge view reveals whether ringing came from the source or from the operator you stacked afterward.
Signals hiding in plain sight
Edge filters measure change. Flat regions stay dark. Ramps and borders light up because neighboring pixels disagree. Threshold turns those magnitudes into a binary mask: above the line reads as white, below reads as black.
- Sobel blends horizontal and vertical 3×3 kernels, so large steps read louder than gentle ramps.
- Laplacian looks for second derivatives, so thin double responses appear around lines.
- Roberts uses 2×2 diagonals, so tiny shifts matter on noisy phone captures.
| Operator | Bias | Good when |
|---|---|---|
| Sobel | Balanced gradients | General outlines, first pass |
| Prewitt | Similar to Sobel, slightly softer | Noisy scans where Sobel feels twitchy |
| Roberts | Diagonal emphasis | Small images, quick tests |
| Laplacian | Second derivative | Finding zero-crossing style detail |
| Canny (preview) | Sobel magnitude scaled | Quick look, not full hysteresis |
Native apps do not own this workflow
Desktop suites still win for batch jobs, layered masks, and calibrated print output. This tab wins when you need an answer in thirty seconds on hardware you do not control.
If you need a hard black-white cut without gradients, threshold is the blunt sibling: one luminance split instead of gradient magnitude. If you need EXIF sanity before trusting edges, open metadata viewer on the same file first.
For composition fixes before you judge lines, cropper removes dead border pixels that otherwise fake edges along the frame.
What we refuse to simulate
The Canny option here scales Sobel output. Full Canny needs Gaussian blur, non-maximum suppression, and hysteresis linking. We label the menu honestly so you do not cite this page as a reference implementation.
Very large images may stutter because every slider move recomputes the whole raster on the CPU. Downscale in another tool first when files exceed a few megapixels.
JPEG blocks and chroma subsampling create false edges along flat skies. Denoise or re-export from a cleaner source when lines appear where none exist on the subject.
Quick facts you might not expect
Inverting the map swaps ink and paper visually, which helps when you paste onto dark slide themes.
The soften toggle nudges mid-gray responses toward full black or white. Use sparingly; aggressive smoothing erases hairline detail you still wanted.
Roberts ignores a one-pixel border on the bottom-right because the 2×2 kernel never completes there. Expect slightly thinner edges near those edges compared with 3×3 operators.
. Operator math follows standard textbook kernels; UI behavior matches the canvases shown above.
