DualSigVL: confidence-aware expert fusion for offline signature verification based on dualview multimodal large language models

Offline signatures serve as a fundamental basis for financial instrument verification and legal document authentication. Since signature authenticity directly impacts property security and the establishment of legal facts, automated verification systems require high discriminative accuracy. However, existing approaches, predominantly based on Siamese networks and metric learning, suffer from clear performance bottlenecks when confronted with forgeries, limiting their applicability in real-world scenarios. To address this limitation, we propose DualSigVL, which reformulates offline signature verification within an instruction-driven Visual Question Answering (VQA) paradigm using the multimodal large language model Qwen3-VL. DualSigVL employs a complementary dual-expert architecture: the raw-image expert captures appearance-level semantic cues, while the multi-feature expert encodes geometric and structural characteristics. Both experts share a frozen Qwen3-VL backbone and achieve functional specialization via independent Low-Rank Adaptation (LoRA) parameters. To quantify the decision reliability of each expert, a confidence extraction strategy based on the first-token logit is introduced, transforming generative outputs into continuous discriminative confidence scores. Building upon this, the Dual-Expert Confidence-Aware Mixture of Experts (DECA-MoE) module performs confidence-aware fusion via score scaling and dynamic gating in the logit space, enabling sample-adaptive decision-making. Extensive experiments on the CEDAR, BHSig-B, BHSig-H, and ChiSig datasets show that DualSigVL achieves accuracies of 100.00%, 98.33%, 98.70%, and 96.38%, respectively, demonstrating competitive or improved performance under the adopted evaluation protocol. The source code is publicly available at https://github.com/BEFINE-code/DualsigVL .