Densecap-Net: A Hybrid DenseNet – 121 and Capsule Network Architecture for Offline Signature Verification
International journal of intelligent engineering and systems · 2025
Signature verification constitutes a critical component in biometric-based identity authentication systems, particularly for safeguarding the integrity of official, financial, and legal documentation.Offline signature verification (OSV) refers to the analysis of static, two-dimensional representations of handwritten signatures, typically acquired via scanning or image capture, without access to dynamic behavioral traits such as pen pressure, tilt, velocity, or stroke sequencing.Conventional machine learning and handcrafted feature-based approaches often exhibit suboptimal performance due to high intra-class variability-resulting from natural variations in an individual's signature-and inter-class similarity, especially in the presence of skilled or simulated forgeries.These limitations impede generalization and increase false acceptance/rejection rates in real-world scenarios.To overcome these challenges, this study proposes a hybrid deep learning architecture integrating Dense Convolutional Network (DenseNet-121) with a Capsule Network (CapsNet) for enhanced discriminative feature extraction and spatial hierarchy preservation.DenseNet-121 facilitates efficient gradient propagation and feature reuse by employing dense connectivity across layers, enabling robust hierarchical representation learning.The Capsule Network complements this by maintaining part-whole relationships through dynamic routing mechanisms, thereby retaining the orientation and spatial configuration of signature strokes-attributes often lost in traditional CNN-based models.The proposed model was trained and evaluated on a publicly available offline signature dataset from Kaggle, comprising both genuine and forged instances.Experimental results demonstrate that the hybrid model significantly outperforms traditional classifiers and several contemporary deep learning baselines, achieving an accuracy of 98.80%, precision of 98.42%, recall of 99.20%, and an F1-score of 98.81%.These findings underscore the model's robustness, discriminative capability, and generalization power in complex OSV tasks.The synergistic integration of DenseNet and Capsule Network not only enhances the representation of subtle signature dynamics but also ensures scalability and adaptability across diverse signature styles and forgery types.