Temporal Residual Networks basedbeta-elliptic model and multi-head attentionfor online handwriting signature verification

Abstract In this work, we proposed a new system for online handwriting sig-nature verification based on beta elliptic modeling and TemporalResidual Neural Networks based Multi-Head Attention model. Thebeta-elliptic modeling was applied to divide the handwriting signaturetrajectory into strokes by inspecting the extremum velocity instantsand extract their dynamic and geometric proprieties. In the verifica-tion process, residual networks based temporal convolution was devel-oped to deals with sequential input data. Furthermore, we appliedMulti head attention model to focus on a few particular aspectsat a time and carry out an efficient and sequential data process-ing . The experiments were done on two public databases SVC-2004and SCUT-MMSIG, and we achieved the state-of-the-art performanceswith an Equal Error Rates equals to 0.114 and 0. 133 respectively.