Multi-scale Attention-based Individual Character Network for Handwritten Signature Verification
Ziyun Zeng · 2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA) · 2022
Offline handwritten signature verification dictates a neural network to learn from training signature features and tell whether a tested signature is genuine or not. One common idea is that the author first input a reference signature and a tested signature of their original and gray-inverse version. After a couple of convolution and attention, each two of them are concatenated together. With the strategy that all the outcomes should be same, the model is trained to focus more on the signature stroke features than on color. In this paper, the author proposes a novel Siamese network to extract signature stroke features. The database the author uses contains a large scale and challenging signature from reference writers and sophisticated calligraphers. To improve the performance of the model, the author pre-processes all input signatures. The network reaches an accuracy of 82% on SigComp2011 dataset.