---
kind: concept
slug: sobel-edge-detection
title: Sobel-style edge detection
summary: >-
  Sobel-style edge detection marks pixels that differ from neighbours on both axes. One paper uses two thresholded variants, a shadow map for dark grooves and a light map for bright ridges, with a box-count noise filter, to make binary edge maps of finger-knuckle prints. Single-paper evidence; author-reported identification results only.
tags: [edge-detection, image-preprocessing, biometrics, binary-images]
papers: ["2605.17673"]
related: [canny-edge-detection, chamfer-distance, biometric-identification]
status: draft
updated_at: 2026-11-12
---

## Overview

The problem is to find where image brightness changes sharply, so that a
structure such as a crease or outline becomes a set of marked pixels. The
shared idea in this card is to compare each pixel with neighbours at a fixed
distance `d` and mark it when the difference exceeds a threshold `t` on both
the vertical and the horizontal axis [arxiv:2605.17673 §III.A] [eq. 1] [eq. 2].
The paper calls the classical Sobel operator the starting point and adapts it
with separate thresholds for two polarities [arxiv:2605.17673 §III.A]. Its
output is binary, which keeps storage and comparison cheap [arxiv:2605.17673 §I].

## Variants

Only one paper has been compiled for this concept so far, so the variants below
are the two polarities of that one method, not a survey.

### Shadow map (dark grooves)

A pixel is marked when it is darker than both of its neighbours on each axis
by more than `t` [arxiv:2605.17673 §III.A] [eq. 1] [eq. 2]. The paper uses
`t = 8` and `d = 4` for its finger images [arxiv:2605.17673 §III.B]. Card:
[Finger-knuckle prints recognition based on a Sobel filter](/papers/2605.17673).

### Light map (bright ridges)

The mirror rule: a pixel is marked when it is brighter than both neighbours on
each axis by more than `t` [arxiv:2605.17673 §III.A] [eq. 3] [eq. 4]. The paper
uses `t = 5` and `d = 4` [arxiv:2605.17673 §III.B]. Card:
[Finger-knuckle prints recognition based on a Sobel filter](/papers/2605.17673).

### Box-count noise filter

Not an edge rule itself. A pixel stays marked only if the window around it
holds enough marked pixels. The window half-widths start at 10 (shadow) or 5
(light) and shrink to 1 [arxiv:2605.17673 §III.B] [eq. 5]. Card:
[Finger-knuckle prints recognition based on a Sobel filter](/papers/2605.17673).

## When to use which

The card supports only one claim about choice. The paper says the classical
Sobel operator uses a distance `d` close to 1, while its own method uses
`d = 4` on finger images [arxiv:2605.17673 §III.A] [arxiv:2605.17673 §III.B].
A larger `d` looks at coarser structure. The paper does not test other values
of `d`, so this card cannot say which value suits other images.

| Situation | Use | Why |
|---|---|---|
| Binary edge maps for compact storage and fast comparison, on images similar to finger-knuckle prints | The two-polarity rule with a box-count filter, as in [arxiv:2605.17673 §III] | The paper's stated aim is a fast method with small binary outputs [arxiv:2605.17673 §I] [§V] |
| Grey-level gradient magnitude for general edge detection | Not covered by this card | The paper does not compare against gradient-magnitude edge detectors. See [canny-edge-detection](/concepts/canny-edge-detection) |
| Need a result you can trust without re-running it | Re-measure on your own data first | All results are author-reported on one dataset [arxiv:2605.17673 §IV] |

## Implementation sketch

Python-flavoured pseudocode of the shadow rule and the box-count filter. Border
handling is not stated in the paper; the sketch skips out-of-range neighbours.

```pseudocode
# Shadow edge rule  [arxiv:2605.17673 §III.A] [eq. 1] [eq. 2]
# Input:  grey image P (H, W); threshold t; distance d
# Output: binary map SS (H, W)

def shadow_edges(P, t, d):
    out = zeros(H, W)
    for i, j in interior_pixels(P, d):
        # Darker than both neighbours on the vertical axis by more than t
        s1 = (P[i-d, j] - P[i, j] > t) and (P[i+d, j] - P[i, j] > t)
        # Darker than both neighbours on the horizontal axis
        s2 = (P[i, j-d] - P[i, j] > t) and (P[i, j+d] - P[i, j] > t)
        # Mark only when both axes hold  [eq. 1]
        out[i, j] = 1 if (s1 and s2) else 0
    return out
```

```pseudocode
# Box-count noise filter  [arxiv:2605.17673 §III.B] [eq. 5]
# Input:  binary map B (H, W); threshold t; half-widths a_x, a_y
# Output: cleaned binary map (H, W)

def denoise(B, t, a_x, a_y):
    out = zeros(H, W)
    for i, j in pixels(B):
        # Count marked pixels in the window centred on (i, j)  [eq. 5]
        window_sum = sum(B[i-a_y : i+a_y+1, j-a_x : j+a_x+1])
        # The paper's text says "at least t"; eq. 5 says "> t"  [§III.B]
        out[i, j] = 1 if window_sum > t else 0
    return out
    # The paper shrinks a_x and a_y to 1 by repeating this step; whether each
    # pass reads the previous output or the original map is not stated.
```

The light map uses the mirror inequalities of [eq. 3] and [eq. 4]. Crop a
220×110 region of interest from each binary map; the paper takes the ROI from
its dataset source [arxiv:2605.17673 §III.C]. A distance between two binary
crops, such as mean absolute difference or clipped Chamfer
([chamfer-distance](/concepts/chamfer-distance)), is then used for matching
[arxiv:2605.17673 §III.D].

## Contradictions

None known across papers; only one paper is compiled here. Points inside that
paper that a reader must decide for themselves:

- The noise rule uses a strict "greater than t" in [eq. 5], while the text says
  "at least t" [arxiv:2605.17673 §III.B]. The two differ at exactly `t`.
- The paper does not say whether each noise-reduction pass uses the previous
  output [arxiv:2605.17673 §III.B].
- The paper does not say how the shadow and light scores are combined into one
  match score [arxiv:2605.17673 §III.D].
- The paper says the ROI comes from the dataset authors and also that it is
  extracted from the binary maps. It is not clear which one its results use
  [arxiv:2605.17673 §III] [§III.C].

## Papers

- [A simple approach for biometrics: Finger-knuckle prints recognition based on a Sobel filter and similarity measures](/papers/2605.17673): two thresholded Sobel-type edge maps, a box-count filter and binary matching for finger-knuckle prints.
