💻 computer science

A simple approach for biometrics: Finger-knuckle prints recognition based on a Sobel filter and similarity measures

This paper proposes a fast and efficient finger-knuckle print recognition method that utilizes a Sobel filter and noise reduction to generate binary images for comparison via similarity measures, achieving a true positive rate of up to 17.02% on a large dataset.

Original authors: E. O. Rodrigues, T. M. Porcino, Aura Conci, Aristofanes C. Silva

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: E. O. Rodrigues, T. M. Porcino, Aura Conci, Aristofanes C. Silva

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a unique "fingerprint" not on the tip of your finger, but on the knuckle (the joint) where your finger bends. Just like your fingerprint, the skin on your knuckle has a specific pattern of lines and valleys that no one else has. This paper proposes a simple, fast, and clever way to identify people using these knuckle patterns.

Here is the breakdown of their method, explained with everyday analogies:

The Goal: A Simple ID Check

Most security systems are like complex, high-tech vaults that take a long time to open. The authors wanted to build a "bicycle lock" version of biometric security: simple, fast, and effective. They wanted to see if they could identify a person just by looking at a photo of their knuckle, without needing heavy-duty computer processing.

Step 1: The "Shadow and Light" Flashlight (Edge Detection)

When you take a photo of a knuckle, it's just a flat image. To find the unique lines, the researchers used a digital tool called the Sobel filter.

Think of this like shining two different flashlights on a textured wall:

The "Shadow" Light: This shines from one side, making the deep valleys cast long shadows.
The "Light" Light: This shines from the other side, highlighting the ridges that catch the light.

By doing this, the computer doesn't just see a photo; it sees a map of the "valleys" and "ridges" of the skin. The result is a black-and-white image where the important lines are white, and the background is black.

Step 2: The "Dust Broom" (Noise Reduction)

When you use those digital flashlights, sometimes you get "static" or "dust"—tiny, random white dots that aren't part of the actual knuckle pattern. These are errors.

The researchers used a "noise reduction" algorithm, which acts like a dust broom. It looks at every single dot on the image. If a white dot is standing all alone (like a piece of dust), the broom sweeps it away. If a group of white dots is clustered together (like a real line on the knuckle), the broom leaves them alone. This cleans up the image so the computer isn't confused by random specks.

Step 3: The "Cutout" (ROI Extraction)

After cleaning the image, they don't need the whole photo. They just need the specific part with the knuckle. They cut out a rectangular "sticker" (called a Region of Interest) that contains only the knuckle pattern, discarding the rest of the finger.

Step 4: The "Matching Game" (Similarity Measures)

Now, the system has a clean, black-and-white "sticker" of your knuckle. To see if it matches someone in the database, it has to compare your sticker against thousands of other stickers.

The paper tested three different ways to play this matching game:

The "Pixel-by-Pixel" Count (Mean Absolute): This is like checking if two puzzle pieces are exactly the same. If even one tiny dot is in the wrong place, it counts as a mismatch. The authors found this was too strict; it didn't care about the overall shape, just the exact position.
The "Worst-Case" Scenario (Hausdorff Distance): This looks for the single biggest mistake. Imagine two shapes are 99% identical, but one has a tiny speck far away. This method would say, "They are totally different because of that one speck." It's very sensitive to errors.
The "Average Neighbor" (Chamfer Distance): This was the winner. Imagine you are trying to match two shapes. Instead of demanding they be identical, this method asks: "For every point on your shape, how close is the nearest point on my shape?" It averages these distances. It's forgiving of small errors but still recognizes the overall pattern.

The Results: How Well Did It Work?

The researchers tested this on a database of 147 people, each with 12 different photos of their knuckles.

The Score: When the system had to guess who a person was from a single photo, it got it right 17.02% of the time.
Why that number is good: If the system were just guessing randomly, it would only get it right about 0.68% of the time. So, their method was roughly 25 times better than random guessing.
The "Accuracy" Trap: The paper notes that the system had a very high "accuracy" (99.99%), but this is a bit of a trick. Because there are so many people who don't match (true negatives), the overall percentage looks huge. The real test is the True Positive Rate (getting the right person), which was the 17.02% mentioned above.

The Conclusion

The authors conclude that their method is simple, fast, and robust.

Because they turned the images into simple black-and-white maps, the computer doesn't have to do heavy math.
It can run on very basic computers.
While 17% might sound low compared to complex systems, the authors point out that if the database were smaller (like a small office of 20 people) or if they scanned the person twice, the success rate would go much higher.

In short, they proved you don't need a supercomputer to recognize a knuckle; you just need a clever way to look at the shadows and lines, clean up the dust, and use a forgiving matching game.

Technical Summary: A Simple Approach for Finger-Knuckle Prints Recognition Based on a Sobel Filter and Similarity Measures

Problem Statement
Biometric systems are widely used for security and individual recognition, yet they often suffer from flaws due to the discrete nature of sampled data, which can suppress recognition quality. While various methodologies exist for finger-knuckle print (FKP) recognition, many are computationally complex and time-consuming. The objective of this work is to propose a novel, simple, and robust methodology for FKP recognition that relies on fundamental computer vision principles to achieve efficient processing and storage without sacrificing the potential for reliable recognition.

Methodology
The proposed approach utilizes the Finger-Knuckle Prints (FKP) dataset provided by Zhang et al., which contains images from 147 volunteers. The methodology follows a streamlined pipeline consisting of preprocessing, region of interest (ROI) extraction, and similarity comparison:

Edge Detection via Adapted Sobel Filters:
Instead of a standard edge detection, the authors apply two distinct adaptations of the Sobel operator to the input images:
- Shadow Sobel: Detects valleys filled with shadows.
- Light Sobel: Detects valleys that are enlightened.
  These filters utilize specific threshold ( $t$ ) and distance ( $d$ ) parameters to generate binary images. The authors evaluated fusing these two images but rejected the approach due to significant noise increase and information loss.
Noise Reduction:
The binary images generated by the Sobel filters contain inherent noise. To address this, a "box-blur-like" noise reduction algorithm is applied. This algorithm evaluates a squared neighborhood around each pixel; if the neighborhood contains at least a threshold number of "painted" pixels, the central pixel is retained. Otherwise, it is erased. This process is iteratively applied with decreasing neighborhood sizes to eliminate isolated noise pixels while preserving the structural lines of the finger knuckles.
ROI Extraction:
Following preprocessing, a Region of Interest (ROI) of 220 x 110 pixels is extracted from the center of the knuckle area. This step can be performed either before or after preprocessing, though the authors performed it after. Each finger yields two ROIs: one from the shadow filter and one from the light filter.
Similarity Measures:
To compare an input finger against the database, the system compares the shadow and light ROIs of the input against the corresponding ROIs of all stored samples. Three similarity measures were evaluated:
- Mean Absolute (MA) / Hamming Distance: Calculates the sum of absolute differences between pixel values.
- Hausdorff Distance (HD): Measures the worst-case difference between two sets of points. The authors noted its sensitivity to outliers, which the noise reduction step aims to mitigate.
- Chamfer Distance (CD): Calculates the mean distance between each point in one set and its closest point in the other, utilizing a threshold to reduce the impact of outliers and missing edges.

Key Contributions

Simplicity and Speed: The paper introduces a methodology based on exceptionally fast operations (Sobel filtering and binary image processing) that produce sparse matrices, making them highly efficient for storage and processing on low-capable systems.
Dual-Filter Approach: The use of separate Shadow and Light Sobel filters to capture different aspects of the knuckle texture without fusing them, thereby avoiding noise accumulation.
Comparative Evaluation: A systematic evaluation of three distinct similarity measures (MA, HD, and CD) within the context of FKP recognition, highlighting the trade-offs between exact matching and similarity tolerance.

Results
The system was tested on a large dataset involving 147 volunteers with 12 images each (totaling 1763 comparisons per candidate).

Recognition Rates: The Chamfer Distance achieved the highest True Positive (TP) rate of 17.02%. In comparison, the Hausdorff Distance achieved 2.60%, and the Mean Absolute measure achieved 0.78%.
Accuracy: All methods achieved high accuracy rates (above 99.88%), though the authors note that accuracy is heavily influenced by the large number of True Negatives in the dataset.
Processing Time: The Mean Absolute measure was the fastest. The Chamfer Distance was approximately 105 times slower than the Mean Absolute measure, while the Hausdorff Distance was roughly 223 times slower.
Per-Finger Performance: Recognition rates varied by finger, with the Left Index achieving the highest TP rate (30.77%) and the Right Index the lowest (18.21%).
Statistical Significance: The authors note that a random recognition rate would be approximately 0.68%. Achieving 17.02% is considered "very expressive" in this context.

Significance and Claims
The paper claims that its primary significance lies in the simplicity and robustness of the approach. By constraining the methodology to basic computer vision techniques (Sobel operators and binary processing), the system is highly optimized for low-resource environments. The authors emphasize that while the 17.02% TP rate is modest compared to some complex state-of-the-art methods, it is achieved with significantly less computational overhead.

The paper modestly suggests that the recognition rate could be improved by:

Using multiple samples per user (scanning a finger more than once).
Testing grey-image oriented similarity measures (though acknowledging this might require slower registration processes).
Applying machine learning algorithms in future work.

The authors conclude that for small datasets (e.g., up to 20 users), the approach could potentially reach maximum TP rates, and the use of binary images allows for substantial optimization in both storage and processing time.