Most of the current electronic signature handwriting recognition models mainly focus on time series features and ignore the correlation between dimensions, and the feature modeling ability of time series features is insufficient. To solve this problem, an electronic signature handwriting recognition model based on double-branch Transformer was proposed to realize high-precision recognition of electronic signature handwriting. Firstly, a dual-branch feature extraction structure of Transformer was designed to fully capture temporal features and channel features.Secondly, in order to realize multi-scale temporal feature learning, a multi-scale feature extractor was introduced in the temporal feature branch to enhance the encoder’s fine-grained learning of temporal features, and a feature recalibration module was introduced in the encoder to strengthen the learning of key points in the signature time sequence.In addition, a channel attention module encoder with a dual-view pooling mechanism was proposed in the channel feature branch to enhance the correlation between features of different dimensions.Finally, a dual-branch feature fusion module was designed to effectively fuse complementary temporal features and channel features using a hierarchical design of local enhancement first and global fusion later. The performance evaluation was carried out on the public datasets MSDS and MCTY. The EER(equal error rate) of the simulation scene 1 vs. 1 achieves 2.54% and 2.61% respectively, and the EER of the simulation scene 4 vs. 1 achieves 1.64% and 1.87% respectively. The experimental results show that the proposed method has higher recognition accuracy and has certain advantages in complex scenes.
然后对不同尺度的特征使用多感受野融合模块(multi-receptive field fusion, MRFF)实现并行分支特征的自适应融合,对各个分支提取的特征Fi使用全局池化提取特征gi,再对各分支使用平均池化得到特征gavg,基于全局特征自适应评估各个卷积核分支的重要性权重αi,对各卷积分支特征Fi进行转置后得到FTi,对FTi使用评估重要性权重αi进行元素级乘法加权,将各个分支特征进行连接得到Fcontact,最后融合得到最终的Ffused,获得更全面的特征表示。结构如图2(a)所示,其中特征计算、分支重要性权重计算和融合特征的计算如式(1)~式(4)所示。
基于双分支Transformer的电子签名笔迹识别模型在MSDS和MCTY数据集上进行实验,如图4所示,本文模型ROC曲线下面积(area under the curve,AUC)为0.996 5,说明模型能在保持极低假阳性率,同时成功捕获大部分真阳性样本,表明模型具有优秀的分类能力。并与Sig2Vec[24]、TA-RNNs[25]、DsDTW[26]、BiLSTM[27]、DWSCNN[28]等近几年主流模型对比,在MSDS数据集和MCTY数据集上均进行1对1和4对1场景,特别的是在MSDS和MCTY数据集上进行跨阶段数据实验,验证模型的泛化能力,结果如表2所示。通过实验结果可以得出本文模型在MSDS数据集中,单阶段的1对1场景和4对1场景均取得最优或者次优的结果,跨阶段的1对1场景中模型EER实现2.54%,相较于次优的TA-RNNs(9.04%)有显著改善,4对1场景中所提出模型EER达到1.64%,同样大幅优于所有对比模型。这表明本文提出的模型在处理不同阶段的签名样本时具有更强的适应性,这对实际应用场景至关重要。
BhavaniS D, BharathiR K. A multi-dimensional review on handwritten signature verification: strengths and gaps[J]. Multimedia Tools and Applications, 2024, 83(1): 2853-2894.
FanZiyan, WangYuanfeng. Scientific approaches to handwrittenelectronic signature authentication[J]. Forensic Science and Technology, 2025, 50(4): 405-412.
[8]
LiuR, XinY. Online handwritten signature verification method based on unified feature correlation coefficient between signatures[J]. Sensors, 2023, 23(23): 9341.
[9]
SaleemM, KovariB. Online signature verification using signaturedown-sampling and signer-dependent sampling frequency[J]. Neural Computing and Applications, 2025, 37(2): 639-651.
[10]
ParzialeA, DiazM, FerrerM A, et al. SM-DTW: stability modulated dynamic time warping for signature verification[J]. Pattern Recognition Letters, 2019, 121: 113-122.
[11]
SekharV C, MukherjeeP, GuruD S, et al. Online signature verification based on writer specific feature selection and fuzzy simil-arity measure[J]. Pattern Recognition Letters, 2023, 174: 40-47.
[12]
LaiS, JinL. Recurrent adaptation networks for online signatureverification[J]. IEEE Transactions on Information Forensics and Security, 2019, 14(6): 1624-1637.
[13]
TolosanaR, Vera-RodriguezR, FierrezJ, et al. Exploring recur-rent neural networks for on-line handwritten signature biometrics[J]. IEEE Access, 2018, 6: 5128-5138.
[14]
WenQ, ZhouT, ZhangC, et al. Transformers in time series: a survey[C]// Proceedings of the Thirty-second International Joint Conference on Artificial Intelligence. Macao: International JointConferences on Artificial Intelligence Organization, 2023: 6778-6786.
[15]
VoruguntiC S, GautamA, PulabaigariV. OSVConTramer: a hy-brid CNN and transformer based online signature verification[C]// IEEE International Joint Conference on Biometrics (IJCB). Ljubljana: IEEE, 2023: 1-10.
[16]
MelziP, TolosanaR, Vera-RodriguezR, et al. Exploring transfo-rmers for on-line handwritten signature verification[J]. arXiv Preprint, arXiv: 2307.01663, 2023.
RenLingling, WangLi, HuangXuewen, et al. Four kinds of EEGsignal decoding based on attention and multi-scale[J]. Science Technology and Engineering, 2022, 22(34): 15180-15187.
[19]
HouS, SunS, YinT, et al. AMDCnet: attention-gate-based multi-scale decomposition and collaboration network for long-term time series forecasting[J]. Frontiers in Artificial Intelligence, 2025, 8(8): 1607232.
[20]
SaleemM, SzucsC, KovariB. Systematic evaluation of pre-processing approaches in online signature verification[J]. Intelligent Decision Technologies, 2023, 17(3): 655-672.
[21]
WangZ, WangH, JiaT, et al. DBConformer: dual-branch convolutional transformer for EEG decoding[J]. IEEE Journal of Biomedical and Health Informatics, 2025. DOI: 10.1109/JBHI.2025.3622725.
ZhuZhenfei, GeDongyuan, YaoXifan, et al. Environmental sound classification model combining Swin Transformer and CNN[J]. Science Technology and Engineering, 2024, 24(28): 12259-12267.
[24]
ZhangP, JiangJ, LiuY, et al. MSDS: a large-scale chinese signature and token digit string dataset for handwriting verification[J]. Advances in Neural Information Processing Systems, 2022, 35: 36507-36519.
[25]
Ortega-GarciaJ, Fierrez-AguilarJ, SimonD, et al. MCYT base-line corpus: a bimodal biometric database[J]. IEEE Proceedings-Vision, Image and Signal Processing, 2003, 150(6): 395-401.
[26]
WangM, DengW. Deep face recognition: a survey[J]. Neurocomputing, 2021, 429: 215-244.
LaiS, JinL, ZhuY, et al. SynSig2Vec: forgery-free learning of dynamic signature representations by sigma lognormal-based syn-thesis and 1D CNN[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(10): 6472-6485.
[29]
TolosanaR, Vera-RodriguezR, FierrezJ, et al. BioTouchPass2: touchscreen password biometrics using time-aligned recurrent neu-ral networks[J]. IEEE Transactions on Information Forensics and Security, 2020, 15: 2616-2628.
[30]
JiangJ, LaiS, JinL, et al. Dsdtw: local representation learning with deep soft-dtw for dynamic signature verification[J]. IEEE Transactions on Information Forensics and Security, 2022, 17: 2198-2212.
[31]
LiC, ZhangX, LinF, et al. A stroke-based RNN for writer-in-dependent online signature verification[C]// International Confe-rence on Document Analysis and Recognition(ICDAR). Sydney: IEEE, 2019: 526-532.
[32]
ZhengJ, ChenZ, HuangL, et al. DWSCNN online signature ve-rification algorithm based on CAE-MV feature dimensionality re-duction[J]. IEEE Access, 2024, 12: 22144-22157.