Multi-scale residual based siamese neural network for writer-independent online signature verification

How to extract effective features from a small number of signature samples provided by new registered users and achieve high accuracy authentication is a challenge for researchers in the field of online signature verification. In response to this challenge, we propose a novel writer-independent online signature verification method using the combin ation of one dimensional multi-scale residual based Siamese neural network (1D-MRSNet) and adaptive boosting softmax (ABSoftmax) classification. First of all, we propose a channel attention mechanism based Siamese neural network (CASNet) framework. The proposed CASNet framework can adaptively learn the weights of different signature feature sequences, which makes the network pay more attention to the information of important feature sequences. Since the proposed CASNet framework does not need to train the signature samples registered by new users, it effectively alleviated the problem of insufficient signature samples in practical applications. Next, a multi-scale residual attention mechanism (MSRA) block is proposed to automatically extract the multi-scale features of the signature, so as to improve the representation learning ability of the network. Then, the adaptive boosting (AdaBoost) algorithm is used to construct ABSoftmax classifier to achieve the integrated decision of online signature verification, thereby improving the accuracy of online signature verification. Finally, when using a reference sample, the proposed method has achieved 6.57% equal error rate (EER) and 11.74% EER in MCYT-100 database and SVC 2004-task2 database, respectively. When using five reference samples, the proposed has method achieved 1.38% EER and 2.33% EER in two databases respectively. The results illustrate the proposed method reached a state-of-the art level.