Paper · 2026-05-17 · cs.CV
A simple approach for biometrics: Finger-knuckle prints recognition based on a Sobel filter and similarity measures
Finger-knuckle print matching with two Sobel-type edge maps (shadow and light), a box-count noise filter, a 220x110 ROI and binary comparison by mean absolute difference, Hausdorff or clipped Chamfer distance. Author-reported: 17.02% true-positive recognition with Chamfer distance on the 147-volunteer FKP dataset, against 15.15% for a local-feature method cited from another paper.
Chamfer distanceSobel-style edge detectionBiometric identificationbiometricsfinger-knuckle-printedge-detectionsimilarity-measures
AI-compiled by claude-haiku-5-5 on 2026-11-12, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
In plain words
A finger's middle joint has a pattern of fine creases, and that pattern can identify a person, much like a fingerprint. This paper matches people by that pattern using deliberately simple image processing. It finds the creases twice, once for the dark grooves and once for the bright ridges, turns each result into a black-and-white picture, removes speckle, and crops a fixed box around the joint. Two cropped pictures are then compared with a plain distance: how far apart their marked pixels are.
To identify someone, the incoming finger is compared with every finger in a database, and the closest match wins. The authors report that the best of three distances gets the right person in about 17% of comparisons. They present this as useful because the method is fast and its outputs are small binary images.
TL;DR
The method turns each finger image into two binary edge maps with Sobel-type filters, a shadow map and a light map, then removes isolated pixels with a box-count filter §III.A §III.B. It crops a 220×110 region of interest §III.C and compares the binary crops with mean absolute difference, Hausdorff distance or a clipped Chamfer distance §III.D. Author-reported, Chamfer distance gives 17.02% true positives on the authors' FKP setup, with the other two distances far lower tab. I. All results come from one dataset, and the comparison with other works mixes metrics §IV tab. IV.
Problem
Finger-knuckle print recognition is a biometric task, and the authors note that the existing methods they cite are complicated or slow §II. They want a simple method that is fast and produces compact binary images for storage and matching §I §V. The matching step depends on a region of interest around the knuckle; the paper uses the ROIs supplied with the dataset and points to the dataset source for how they were obtained §III §III.C.
Method
The method is a fixed pipeline with four stages, shown in fig. 1 §III:
- Input. Take a finger image at 384×288 pixels; the paper gives this resolution for its inputs §III.B.
- Edge maps. Compute two binary maps from the same image: a shadow map that marks grooves darker than their neighbours, and a light map that marks ridges brighter than their neighbours §III.A fig. 2. A pixel is marked only when the condition holds in both axes eq. 1 eq. 3. The offset d sets the neighbour distance, and t sets the contrast needed.
- Noise reduction. Keep a pixel only if the window around it has enough marked pixels eq. 5 §III.B fig. 3. The window half-widths are shrunk step by step until both are 1 §III.B.
- ROI. Crop a 220×110 pixel region around the knuckle from each binary map, using the ROI given by the dataset authors §III.C §III. The paper also says the crop can be done before preprocessing §III.C.
- Compare. Compare the binary ROIs with one similarity measure §III.D. The three measures are the mean absolute difference eq. 6, the Hausdorff distance eq. 7 and the Chamfer distance clipped at τ eq. 8.
- Match. Each finger yields two ROIs (shadow and light). The gallery finger with the smallest distance is taken as the match §III.D §IV.
The edge rules, written with the paper's notation (P is the grey image, P_{i,j} is the pixel at line i and column j, t the contrast threshold, d the neighbour distance, ∧ the boolean and) eq. 1 eq. 2 eq. 3 eq. 4:
SS(P,d,t,i,j) = 1 \text{ if } s_1 \wedge s_2, \text{ else } 0 \quad [\text{eq. 1}]
s_1 = (P_{i-d,j} - P_{i,j} > t) \wedge (P_{i+d,j} - P_{i,j} > t) \quad [\text{eq. 2}]
s_2 = (P_{i,j-d} - P_{i,j} > t) \wedge (P_{i,j+d} - P_{i,j} > t) \quad [\text{eq. 2}]
LS(P,d,t,i,j) = 1 \text{ if } l_1 \wedge l_2, \text{ else } 0 \quad [\text{eq. 3}]
l_1 = (P_{i,j} - P_{i-d,j} > t) \wedge (P_{i,j} - P_{i+d,j} > t) \quad [\text{eq. 4}]
l_2 = (P_{i,j} - P_{i,j-d} > t) \wedge (P_{i,j} - P_{i,j+d} > t) \quad [\text{eq. 4}]
The noise rule keeps a pixel when its window sum is above t eq. 5. The window is centred on the pixel with half-widths a_x and a_y §III.B.
NR(P,t,i,j) = 1 \text{ if } \sum_{y} \sum_{x} P_{y,x} > t, \text{ else } 0 \quad [\text{eq. 5}]
The window sums run over the window centred on (i, j). The PDF's subscripts are hard to read at page resolution, so check them against the PDF before implementing.
The three distances, where P_1 and P_2 are two binary maps and p_1, p_2 are their marked points §III.D:
MA(P_1,P_2) = \frac{1}{hw} \sum_{y=0}^{h} \sum_{x=0}^{w} |P_{1,y,x} - P_{2,y,x}| \quad [\text{eq. 6}]
HD(P_1,P_2) = \max\Big\{ \sup_{p_1 \in P_1} \inf_{p_2 \in P_2} \|p_1 - p_2\|,\ \sup_{p_2 \in P_2} \inf_{p_1 \in P_1} \|p_1 - p_2\| \Big\} \quad [\text{eq. 7}]
CD(P_1,P_2) = \frac{1}{n_1} \sum_{p_{k_1} \in P_1} \max\Big( \min_{p_{k_2} \in P_2} \|p_{k_1} - p_{k_2}\|,\ \tau \Big) \quad [\text{eq. 8}]
The paper's pseudocode, in our own notation, is below. It follows the text; where the paper leaves a choice open, the code says so.
# Finger-knuckle print matching [fig. 1] [§III]
# Input: probe image; gallery of (identity, image) pairs; measure
# Output: identity with the lowest distance
def identify(probe, gallery, measure):
# Shadow and light ROIs of the probe, after noise reduction [§III.A-C]
probe_rois = extract_rois(probe)
best_identity = None
best_distance = infinity
for identity, image in gallery:
# Gallery ROIs by the same steps [§III.A-C]
gallery_rois = extract_rois(image)
# Combining the shadow and light scores is not stated [§III.D]
distance = combine(measure(probe_rois, gallery_rois))
if distance < best_distance:
best_distance = distance
best_identity = identity
return best_identity# Edge maps [§III.A] [eq. 1] [eq. 2]
# Input: grey image P (H, W); threshold t; distance d
# Output: binary shadow map SS (H, W)
def shadow_edges(P, t, d):
# Darker than both neighbours on the vertical axis by more than t [eq. 2]
vertical = (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 [eq. 2]
horizontal = (P[i, j - d] - P[i, j] > t) and (P[i, j + d] - P[i, j] > t)
# Mark the pixel only when both axes hold [eq. 1]
return vertical and horizontal# Noise reduction [§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):
# Window sum around each pixel, centred on it [eq. 5]
for i, j in pixels(B):
window_sum = sum of B over rows i-a_y..i+a_y, cols j-a_x..j+a_x
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 feeds the next is not stated [§III.B]# Chamfer distance, one direction [§III.D] [eq. 8]
# Input: point sets A (n1 points), B; clip threshold tau
# Output: mean clipped distance from A to B
def chamfer(A, B, tau):
total = 0
for p in A:
# Distance to the nearest point of B, clipped at tau [eq. 8]
nearest = min(distance(p, q) for q in B)
total = total + max(nearest, tau)
return total / count(A)Hyperparameters
| Name | Value | Where | Notes |
|---|---|---|---|
| input image size | 384×288 pixels | §III.B | resolution stated for the Fig. 2 images |
| shadow edge threshold t | 8 | §III.B | used to generate the shadow image in fig. 2 |
| shadow edge distance d | 4 | §III.B | classical Sobel would use d close to 1 §III.A |
| light edge threshold t | 5 | §III.B | used to generate the light image in fig. 2 |
| light edge distance d | 4 | §III.B | |
| shadow noise threshold t | 13 | §III.B | chosen empirically |
| light noise threshold t | 7 | §III.B | chosen empirically |
| shadow noise half-widths a_x = a_y | 10 initially, shrunk to 1 | §III.B | window of 21×21 at the start (derived) |
| light noise half-widths a_x = a_y | 5 initially, shrunk to 1 | §III.B | window of 11×11 at the start (derived) |
| ROI size | 220×110 pixels | §III.C | obtained as in the dataset source [3] |
| Chamfer clip τ | 2 | §III.D | |
| matching rule | smallest distance | §III.D §IV | "maximum similarity" is also mentioned |
| dataset | 147 volunteers, 12 sessions, 4 fingers | §III §IV | left/right index and middle; sessions about 25 days apart on average |
| learning or training | not stated | no learning stage is described | |
| hardware, runtime environment | not stated |
Results
Author-reported. Setup: the Zhang et al. finger-knuckle-print dataset §III; 147 volunteers, 12 sessions per finger §III; each candidate gets 1763 comparisons (12×147−1), of which 12 are genuine §IV.
Edge-measure comparison, as the paper reports it tab. I. The labels are the paper's; see the Pitfalls section for why the columns need care.
| Measure | TP % | TN % | FP % | FN % |
|---|---|---|---|---|
| Chamfer distance | 17.0238 | 99.9877 | 82.9762 | 0.0123 |
| Hausdorff distance | 2.6046 | 99.9855 | 97.3953 | 0.0144 |
| Mean absolute difference | 0.7856 | 99.9852 | 99.2113 | 0.0148 |
Accuracy and relative time, with mean absolute difference as the time baseline tab. II:
| Measure | Accuracy % | Time (relative to mean absolute) |
|---|---|---|
| Chamfer distance | 99.9902 | 105.0234 |
| Mean absolute difference | 99.9854 | 1 |
| Hausdorff distance | 99.8887 | 223.557 |
Chamfer distance per finger, with the four fingers evaluated separately tab. III:
| Finger | TP % | FN % | Accuracy % |
|---|---|---|---|
| Left index | 30.7738 | 0.0412 | 99.9176 |
| Left middle | 26.3690 | 0.0438 | 99.9123 |
| Right index | 18.2142 | 0.0487 | 99.9026 |
| Right middle | 26.4285 | 0.0438 | 99.9126 |
Comparison with works the authors cite tab. IV:
| Work | TP % | Accuracy % |
|---|---|---|
| This work | 17.023 | 99.99 |
| Zhang et al. [3] | 15.15 | not given |
| Le-qing [4] (middle finger only, per §II) | not given | 96.91 |
| Guru et al. [6] | not given | 92.24 |
The paper's headline is the 17.02% true-positive rate with Chamfer distance §IV [abstract]. It reports that random matching would give 0.680% (12 out of 1763) §IV. The text also gives a probability of about 1.1×10⁻²⁴ % for recognising every user by chance, without showing how it was computed §IV.
Paper vs code
No official code.
Searched: the paper text (no repository link); the arXiv abstract page (no code link, no comments field); the GitHub repository search API for "finger knuckle sobel", "finger knuckle print", "knuckle sobel Conci" and "Finger-knuckle Sobel filter" (the last returned HTTP 403); and the GitHub user search for the first author's name (no result). The repository hits were student and course projects by unrelated owners and were not used.
Pitfalls
- Table I labels. In every row, TP and FP add to about 100 (for example 17.0238 + 82.9762), and TN and FN add to 100. The FP and FN columns therefore look like complements of TP and TN, not independently measured error rates. Confirm what the columns mean before quoting them. The paper calls TP "the most important index" and notes that TN is large whenever the number of comparisons is huge §IV.
- TP is counted per comparison. The paper's TP is a rate over all probe-gallery comparisons (12 genuine out of 1763 per candidate) §IV. How this maps to a single identification decision is not stated.
- Two scores per finger. Each finger has a shadow ROI and a light ROI. The paper says the shadow and light ROIs of one finger are compared with those of the other, and the pair with the least distance decides the match §III.D. It does not say whether the pairs are matched like with like, or how a shadow distance and a light distance are combined into one score, so an implementer has to choose and should state the choice.
- Strict versus "at least". The noise rule eq. 5 uses a strict "> t". The text says the neighbourhood must hold "at least t pixels" §III.B. The two differ at exactly t.
- Noise iteration. The paper says the window half-widths are decremented until both are 1 and that eq. 5 is "called again" each time §III.B. It does not say whether each pass uses the output of the previous pass or the original image.
- Window sums are hard to read. The summation limits in eq. 5 are small in the PDF. The centred form above is read from the text, so check it against the PDF before coding.
- Mean absolute bounds. The sum in eq. 6 runs to h and w as written, but the normalisation is 1/(hw). Whether the indices run to h−1 and w−1 is not stated. Use the normalisation that makes the measure a mean over pixels.
- Chamfer direction. eq. 8 is one-directional: it averages from P_1 to P_2 only. The paper does not say whether the two directions are averaged, and the one-directional form is not symmetric. The conclusion suggests computing distances only between non-background pixels, but it is not stated whether the experiments did so §V.
- Hausdorff sensitivity. The authors say Hausdorff distance is very sensitive to outliers and that their noise reduction only partly reduces this §IV. Its low TP (2.6046%) should be read with that in mind.
- Fusion failed. Fusing the shadow and light images raised the noise, even after noise reduction, and registration raised processing time without a clear gain §III.B. These variants are not in the results.
- ROI source. The paper says it uses the ROIs provided by the dataset authors §III and also says the ROI is extracted from the binary maps §III.C. It is not clear which ROI the reported results use.
- Fingers pooled or separate. The count of 1763 comparisons fits one finger type per gallery. How the four fingers are combined in Table I is not stated, and Table III reports them separately §IV.
- Comparison with other works. Table IV sets this paper's TP against another paper's TP and its accuracy against other papers' accuracy. The metrics and data splits differ, so the rows are not directly comparable §IV.
- Untested claims. The conclusion says results would improve if each user were scanned at least twice, and that a small dataset (about 20 users with 12 images each) could reach 100% TP §V. Neither was tested in the paper.
- Accuracy is undefined. Table II reports accuracy without defining it, and it does not equal the TN column (for example 99.9902 versus 99.9877 for Chamfer distance).
Lineage
- Builds on: the Sobel operator, which the authors adapt with separate shadow and light thresholds §III.A. Its original reference is cited in the paper but has no card here.
- Builds on: the Chamfer distance, cited to Thayananthan et al. (shape context and Chamfer matching) §III.D. Use the
chamfer-distanceconcept page for background. - Compared against: the Zhang et al. finger-knuckle-print dataset and local-feature method [3] [5], the SURF-based method of Le-qing [4] and the moments and feature-fusion method of Guru et al. [6], all as cited in §II and tab. IV.
- Superseded by: not stated.
Source and license
Adapted from “A simple approach for biometrics: Finger-knuckle prints recognition based on a Sobel filter and similarity measures” by E. O. Rodrigues, T. M. Porcino, Aura Conci, Aristofanes C. Silva (arXiv:2605.17673), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors.