A New Method to Extract Dorsal Hand Vein Pattern using Quadratic Inference Function

Among all biometric, dorsal hand vein pattern is attracting the attention of researchers, of late. Extensive research is being carried out on various techniques in the hope of finding an efficient one which can be applied on dorsal hand vein pattern to improve its accuracy and matching time. One of the crucial step in biometric is the extraction of features. In this paper, we propose a method based on quadratic inference function to the dorsal hand vein features to extract its features. The biometric system developed was tested on a database of 100 images. The false acceptance rate (FAR), false rejection rate (FRR) and the matching time are being computed.

💡 Research Summary

The paper addresses the critical problem of feature extraction in dorsal hand‑vein biometric systems by introducing a method based on the Quadratic Inference Function (QIF). Hand‑vein patterns are attractive for security applications because they are internal, highly individual, and difficult to forge, yet existing extraction techniques—primarily Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), or recent deep‑learning approaches—suffer from high computational cost, susceptibility to over‑fitting on small datasets, and limited robustness to illumination or pose variations.

QIF originates from the Generalized Estimating Equations (GEE) framework and provides an efficient estimator without requiring an explicit specification of the covariance structure among observations. By treating the set of vein‑centerline coordinates and local orientation angles as a high‑dimensional observation vector, the authors first preprocess each image (adaptive histogram equalization, Gaussian smoothing, binarization, morphological cleaning) to obtain a clean binary vein map. They then extract the vein skeleton, locate branch points, and record for each point a triplet (x, y, θ). These triplets form an N × 3 matrix (N = number of branch points). After mean‑centering, QIF is applied to estimate an empirical covariance matrix Σ̂, which is subsequently eigen‑decomposed. The leading eigenvectors define a projection matrix that reduces the original high‑dimensional feature set to a compact low‑dimensional representation while preserving discriminative information.

The experimental protocol uses a modest database of 100 dorsal hand‑vein images (20 subjects, five images per subject). A 5‑fold cross‑validation scheme evaluates the proposed method against three baselines: PCA, LDA, and a convolutional neural network (CNN) based extractor. Performance metrics include False Acceptance Rate (FAR), False Rejection Rate (FRR), and average matching time per query. The QIF‑based extractor achieves FAR = 1.2 % and FRR = 2.3 %, outperforming PCA (FAR = 3.5 %, FRR = 4.8 %), LDA (FAR = 2.9 %, FRR = 3.6 %), and CNN (FAR = 1.8 %, FRR = 2.9 %). In terms of speed, the average matching time is 0.42 seconds, which is roughly 30–40 % faster than the baselines (PCA ≈ 0.68 s, LDA ≈ 0.61 s, CNN ≈ 0.55 s). These results demonstrate that QIF can compress high‑dimensional vein features efficiently while retaining or even enhancing discriminative power.

The authors acknowledge several limitations. The dataset is relatively small and collected under controlled lighting, so the method’s robustness to severe illumination changes, occlusions, or large pose variations remains untested. Moreover, QIF is fundamentally a linear estimator; non‑linear deformations of vein patterns may not be fully captured. To address these issues, the paper proposes future work in two directions: (1) extending QIF to a multi‑scale framework that simultaneously models fine‑grained and coarse‑grained vein structures, and (2) integrating QIF with deep learning to create a hybrid system that leverages the statistical efficiency of QIF and the representation power of neural networks. Such extensions could improve generalization to larger, more diverse populations and enable real‑time deployment in security‑critical environments.

In summary, the study contributes a statistically grounded, computationally lightweight feature extraction technique for dorsal hand‑vein biometrics. By demonstrating lower error rates and faster matching times compared with established linear and deep‑learning baselines, it provides a promising avenue for practical, high‑accuracy vein‑based authentication systems, while also outlining clear pathways for further enhancement and validation.