Hybrid Feature Extraction Based Deep Learning Model for Offline Signature Verification

Offline signature verification serves as a robust authentication method, widely adopted in security applications due to its non-intrusive data collection procedure. Despite the notable advancements in this filed for the past several years, effectively representing signature traits remains challenging. Existing methods rely on handcrafted features or Convolutional Neural Networks (CNN) for visual learning to achieve this goal. Yet, they tend to overlook the intricate nature of signatures, primarily treating them as images while neglecting their multifaceted composition, encompassing continuous curves and strokes. To transcend this limitation, we present a novel hybrid feature extraction based deep learning model for offline signature verification. Considering the implied temporal nature of offline signature, we partition signature images into multiple subregions to extract local features, and then utilize Recurrent Neural Network (RNN) to model the contextual connections between these subregions. For each subregion, we extract corresponding contours features and texture features to exploit the stroke trajectory and static image characteristics of offline signatures. These extracted hybrid features facilitate authentication through following Writer-Dependent classifiers. We evaluate the proposed method on three widely-used datasets, i.e., CEDAR, BHSig-H and BHSig-B. The experimental results indicate that the proposed model can comprehensively represent the input signature, validating its effectiveness. On BHSig-H and BHSig-B, our model improves the performance with equal error rates of 4.61%, 1.49% respectively, compared to 5.65%, 3.96% achieved by previous state-of-the-art methods.