💻 computer science

Pose-Robust Finger Identification Based on Mutual Region Alignment and Minimum Convolution Point Feature

This paper proposes a novel pose-robust finger identification framework that integrates mutual region alignment, minimum convolution point feature learning, and weighted score fusion to effectively address performance degradation caused by longitudinal rotation and non-rigid deformations in contactless finger images, achieving state-of-the-art results across four public datasets.

Original authors: Weili Yang, Dewen Kong, Fujian Feng, Yonghuai Yang, Deng Li, ZhongChi Liu, Dacan Luo

Published 2026-07-29

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Weili Yang, Dewen Kong, Fujian Feng, Yonghuai Yang, Deng Li, ZhongChi Liu, Dacan Luo

Original paper licensed under CC BY 4.0 (https://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 are trying to unlock your phone with your fingerprint. Usually, you just rest your finger flat on the glass, and the scanner reads the ridges. But what if you hold your finger up in the air, like you're pointing at a star, without touching anything? This is called "contactless" scanning, and it's the future of security because it's more hygienic and faster. However, there's a catch: when you hold your finger in the air, it's wobbly. You might twist it slightly, or your finger might bend a little bit. This twisting is like spinning a sausage on a table; the part of the sausage the camera sees changes completely depending on how you spin it. If the camera sees a different slice of your finger every time, the computer gets confused and thinks you are a stranger. This paper tackles that specific problem: how to recognize a finger even when it's twisting, bending, or spinning in the air.

The researchers from Guizhou Minzu University and a power company in China have built a new system to fix this "wobbly finger" problem. They call their method a "pose-robust" framework, which is just a fancy way of saying it works no matter how you hold your finger. They realized that when a finger twists, the camera sees a jumbled mess where the top of the finger in one photo might be the middle in the next. To solve this, they invented a three-step magic trick. First, they use a "mutual region extraction" strategy. Imagine two people holding up maps of the same city, but one map is shifted up and the other is shifted down. The computer looks for the part of the city that appears on both maps and cuts out the rest, aligning them perfectly so they match up. Second, they use a "minimum convolution point feature" to find the unique, bumpy edges of your finger that don't change even if the finger squishes or bends. Finally, they combine this structural "skeleton" of the finger with a "texture" scan (looking at the skin patterns) using a weighted score fusion, kind of like a judge giving points for both the shape of a dancer and the pattern of their costume to decide who wins.

The team tested their new system on four different public databases containing thousands of finger images, including some where people twisted their fingers by up to 80 degrees. The results were impressive. By using their alignment trick, they significantly reduced the error rate. For example, on one dataset called SDUMLA-HMT, their method reduced the error rate (EER) from 4.45% down to just 1.54% when they combined all their techniques. On another dataset, MMCBNU_6000, the error rate dropped to a tiny 0.05%. They found that their method was not only more accurate than older techniques but also faster, taking only 19 milliseconds to process an image compared to 35 milliseconds for a competing method. The paper suggests that by focusing on aligning the overlapping parts of the finger and ignoring the parts that got cut off by the twist, they can make contactless security much more reliable. They conclude that while their system is currently the best at handling twists, future work will try to handle fingers that are also tilted sideways or pitched up and down, making the security system robust against any way you might hold your hand.

Technical Summary: Pose-Robust Finger Identification Based on Mutual Region Alignment and Minimum Convolution Point Feature

Problem Statement

Contactless finger biometric identification systems face significant performance degradation due to longitudinal rotation during image acquisition. This pose variation leads to two primary challenges:

Inconsistent Imaging Regions: Rotation causes the projected imaging plane to shift, resulting in misaligned or missing discriminative regions between the template and probe images.
Non-Rigid Deformation: The inherent flexibility of finger tissues induces deformation under spatial rotation, distorting both finger vein patterns and contactless fingerprint textures.

These factors cause content discrepancies and feature misalignment, severely undermining the stability of feature matching and reducing recognition accuracy. While existing methods address local or global features, they often struggle to simultaneously handle the misalignment caused by rotation and the non-linear deformation of finger tissues.

Methodology

The authors propose a three-stage framework designed to address these specific challenges: Mutual Region Extraction, Minimum Convolution Point Feature Learning with Deformation-Aware Matching, and Weighted Score Fusion.

1. Mutual Region Extraction

To resolve inconsistent imaging areas caused by vertical shifts (longitudinal rotation), the authors propose an adaptive alignment strategy:

Preprocessing: Histogram equalization is applied to enhance contrast and distinguish vein patterns and fingerprint textures.
Vertical Gradient Calculation: The vertical gradient ( $G_y$ ) is computed for each pixel to capture intensity changes along the vertical axis.
Minimum Distance Metric: A template image ( $I_1$ ) is fixed while a probe image ( $I_2$ ) is shifted vertically. The similarity is evaluated by calculating the minimum distance between the vertical gradient images of the overlapping regions.
Alignment: The overlapping region with the minimum distance (highest similarity) is selected as the Mutual Region, effectively eliminating content discrepancies caused by rotation before feature extraction.

2. Minimum Convolution Point Feature Learning & Deformation-Aware Matching

To handle non-rigid deformation and extract robust structural features, the framework introduces a novel feature learning module:

Key-point Detection: Instead of relying solely on linear edges, the method utilizes multi-orientation filters (8 directions: 0° to 157.5°). It computes the minimum convolution response across all orientations for each pixel.
- Rationale: Linear edges may yield zero responses for certain linear filters, whereas non-linear edge patterns (common in complex vein/fingerprint structures) yield non-zero responses across all filters. The minimum convolution value helps distinguish these non-linear points, which are extracted as key-points.
Feature Descriptor: For each key-point, a $3 \times 3$ grid of cells is sampled. The gradient of each cell is quantized into 5 bins (0°–360°). The concatenation of histograms from all 9 cells forms the Minimum Convolution Point Feature.
Deformation-Aware Matching:
- Initial matches are generated using Euclidean distance between point features.
- To remove false matches caused by similar textures, a 2D statistical histogram of offsets (horizontal and vertical) is constructed.
- The method identifies the offset range with the maximum accumulation of matched pairs, assuming correct matches share similar orientation offsets.
- Matches falling outside this optimal range are filtered out. The final similarity score ( $Score_{PM}$ ) is calculated based on the ratio of remaining correct matches to the total number of points.

3. Weighted Score Fusion

Recognizing that structural features (Minimum Convolution Points) excel at handling deformation but may lack fine-grained texture detail, while texture features (Uniform Local Binary Pattern - LBP) capture texture well but are sensitive to deformation, the authors propose a fusion strategy:

LBP Matching: Uniform LBP features are extracted, and matching scores are calculated using Histogram Intersection.
Fusion: The normalized scores from the Minimum Convolution Point matching ( $NormScore_{PM}$ ) and the LBP matching ( $NormScore_{LBP}$ ) are combined using a weighted sum:
$score = w \cdot NormScore_{PM} + (1-w) \cdot NormScore_{LBP}$
where $w$ is a tunable weight parameter.

Key Contributions

Mutual Region Extraction Strategy: A novel method using vertical gradients and minimum distance metrics to adaptively align inconsistent imaging regions, effectively mitigating the impact of longitudinal rotation.
Minimum Convolution Point Feature: A feature learning module that detects non-linear structural patterns via minimum convolution responses across multi-orientation filters, coupled with a deformation-aware matching algorithm to improve robustness against non-rigid finger deformation.
Weighted Score Fusion: A scheme that integrates structural (Minimum Convolution Point) and texture (Uniform LBP) features, leveraging their complementary strengths to enhance identity discriminability and generalizability.

Experimental Results

The proposed method was evaluated on four public datasets: SDUMLA-HMT, MMCBNU_6000, FV-USM, and LFMB-3DFB. Performance was measured using the Equal Error Rate (EER).

Mutual Region Effectiveness: Integrating mutual region extraction significantly reduced EERs across all datasets. For example, on SDUMLA-HMT, the EER for LBP dropped from 12.07% to 4.45%.
Feature Performance: The Minimum Convolution Point feature outperformed traditional statistical features (LBP, HOG, WLD) and classic point-matching methods (fuzzyEnhance-SIFT) on most datasets. On SDUMLA-HMT, it achieved an EER of 2.15%, a 50.7% reduction compared to fuzzyEnhance-SIFT.
Fusion Performance: The weighted fusion strategy achieved the lowest EERs across all datasets, demonstrating the value of combining structural and texture features.
- SDUMLA-HMT: 1.54% (vs. 2.15% for structural alone).
- MMCBNU_6000: 0.05% (vs. 0.52% for structural alone).
- LFMB-3DFB (Fusion of Vein + Fingerprint): 0.12%.
Efficiency: The proposed method has a total processing time of 19ms, which is 45.7% faster than fuzzyEnhance-SIFT (35ms), making it suitable for real-time applications.
State-of-the-Art Comparison: The proposed framework outperformed existing state-of-the-art methods (including Gabor wavelet features, deep representation-based extraction, and dual-sliding window approaches) across all tested datasets.

Significance and Claims

The paper claims that the proposed framework provides a state-of-the-art (SOTA) solution for pose-robust finger identification. Its significance lies in its systematic approach to the core bottlenecks of contactless biometrics:

It addresses imaging region inconsistency through adaptive alignment rather than relying on rigid preprocessing.
It handles non-rigid deformation by modeling non-linear structural features and using a deformation-aware matching algorithm, rather than relying on rigid geometric assumptions.
It overcomes feature insufficiency by fusing structural and texture modalities, ensuring robustness across varying image qualities and pose conditions.

The authors emphasize that the framework's generalization ability stems from its focus on solving fundamental challenges (misalignment and deformation) rather than optimizing for dataset-specific heuristics. Future work is noted to extend this robustness to arbitrary finger poses (lateral tilt, pitch) and explore integration with deep learning models.