Online signature verification based on Siamese networks and LSTM
Fangjun Luan, Jiaqi Qin, Shuai Yuan · 2025
With the rapid development of information technology, online handwritten signature verification has gained significant attention as a biometric technology that offers both security and convenience, particularly in fields such as finance and law. However, existing methods still face challenges in dynamic feature modeling, few-shot learning, and anti-forgery capabilities. This paper proposes an online handwritten signature verification model based on a 1D Convolutional Efficient Channel Attention Long Short-Term Memory Network (1D-ECA-BiLSTM), which effectively improves signature recognition accuracy and robustness by integrating local spatiotemporal features with dynamic sequential modeling. First, the extracted features are screened and dimensionally reduced. A 1D Convolutional Neural Network (1D-CNN) is used to capture the local spatial features of the signature trajectory, while the Efficient Channel Attention (ECA) module adaptively enhances the weights of key channels to optimize feature representation. Second, a Bidirectional Long Short-Term Memory Network (LSTM) is introduced to capture the sequential dynamic characteristics of the signature (such as writing speed and acceleration), addressing the issue of insufficient modeling of long-sequence dependencies in traditional methods. Furthermore, the Siamese Network architecture is employed, with a contrastive loss function to measure signature similarity, enabling efficient learning in few-shot scenarios. The experiment, based on the SVC2004 Task2 standard dataset, shows that the model achieves an accuracy of 97.50% across 40 users, with the Equal Error Rate (EER) reduced to 1.76%. The study confirms that the 1D-ECA-BiLSTM, by combining multimodal feature fusion with attention mechanisms, effectively improves the accuracy of online signature verification.