Offline Signature Verification Using a 2D Attention Encoder-Decoder Network

Verifying an individual's Chinese handwritten signature is a vital biometric technology that is widely used in banking, finance, and legal business. Forged signatures for the purpose of deception endanger these industries' interests. As a result, this paper proposes a network based on 2D Attention to verify the authenticity of the signature. In this paper, we have developed a Chinese handwritten signature dataset (CNSig) and proposed an offline signature verification network (att-OfSVNet) based on 2D Attention. The att-OfSVNet model includes two weight-sharing Encoders and Decoders. The two weight-sharing Encoders receive the inverted genuine and the inverted test signature image, and the Decoder reduces the dimension and concatenates the two extracted feature images. We use 2D Attention to fuse the features extracted by the Encoder and Decoder, which minimizes the information loss in the convolutional layers during the extraction process and enhances the effect of feature extraction by the Encoder. The experimental results show that our att-OfSVNet achieves satisfactory performance on other handwritten signature datasets in three different languages: CEDAR, BHSig-B, and BHSig-H, and it also demonstrates good generalization ability in cross-lingual tests.