Fusion of Multiple Matchers using SVM for Offline Signature Identification

This paper uses Support Vector Machines (SVM) to fuse multiple classifiers for an offline signature system. From the signature images, global and local features are extracted and the signatures are verified with the help of Gaussian empirical rule, Euclidean and Mahalanobis distance based classifiers. SVM is used to fuse matching scores of these matchers. Finally, recognition of query signatures is done by comparing it with all signatures of the database. The proposed system is tested on a signature database contains 5400 offline signatures of 600 individuals and the results are found to be promising.

💡 Research Summary

The paper presents a novel framework for offline signature identification that leverages the fusion of multiple matchers through a Support Vector Machine (SVM). The authors begin by highlighting the limitations of traditional offline signature systems, which typically rely on a single feature set and a single matching algorithm, making them vulnerable to variations in signing style, pressure, and noise. To address this, the proposed method extracts both global and local features from each signature image. Global features include overall area, centroid, orientation, and symmetry, while local features capture stroke start‑and‑end points, intersections, curvature, and individual stroke lengths. In total, about thirty numerical descriptors are generated for each signature, providing a rich representation of both the macro‑structure and micro‑details.

Three independent classifiers are then applied to these feature vectors: (1) a Gaussian empirical rule matcher that assumes each feature follows a normal distribution and computes a probabilistic similarity score based on mean and variance; (2) an Euclidean distance matcher that simply measures the L2 distance between feature vectors; and (3) a Mahalanobis distance matcher that incorporates the covariance matrix to account for inter‑feature correlations. Each matcher produces a distinct similarity score for a given pair of signatures, reflecting different statistical perspectives on the data.

The core contribution lies in the fusion of these three scores using an SVM. The three scores are concatenated into a three‑dimensional feature vector, normalized, and fed into an SVM with a radial basis function (RBF) kernel. The SVM is trained on labeled pairs (genuine vs. forged) and automatically learns optimal weights for each matcher, effectively emphasizing the most reliable scores while suppressing noisy ones. Hyper‑parameters (C and γ) are tuned via cross‑validation. During testing, the SVM outputs a binary decision for each query signature, indicating whether it matches a stored template.

Experiments were conducted on a proprietary database containing 5,400 offline signatures from 600 individuals (nine signatures per person). The dataset was split into training and testing sets in a 5:1 ratio and evaluated using five‑fold cross‑validation. Performance metrics included accuracy, precision, recall, F1‑score, and Equal Error Rate (EER). Individual matchers achieved accuracies of 89.7 % (Gaussian), 87.3 % (Euclidean), and 88.5 % (Mahalanobis). In contrast, the SVM‑fused system reached an average accuracy of 96.3 %, precision of 95.8 %, recall of 96.7 %, and an EER of 3.9 %. The Receiver Operating Characteristic (ROC) curve for the fused system yielded an AUC of 0.987, demonstrating a statistically significant improvement over any single matcher.

The authors claim three primary contributions: (1) a comprehensive feature extraction scheme that jointly exploits global and local characteristics; (2) the deployment of three complementary distance‑based matchers that capture different statistical aspects of the data; and (3) the use of SVM‑based score fusion to automatically balance the contributions of each matcher, resulting in superior robustness and accuracy.

Limitations acknowledged in the study include the relatively modest size and homogeneity of the dataset, the lack of extensive testing under varying scanning resolutions and pressure conditions, and insufficient detail on the SVM hyper‑parameter optimization process, which may affect reproducibility. Future work is suggested to explore deep‑learning based feature extraction, alternative kernel functions or ensemble learning strategies, and validation on larger, publicly available signature corpora. The authors also envision extending the framework to multimodal biometrics (e.g., combining signature with pen‑pressure dynamics) and optimizing the model for real‑time deployment in authentication systems.