SET: a squeeze-and-excitation transformer for offline signature verification

Offline handwritten signature verification, which is widely used in finance, commerce, and criminal forensic identification, plays an essential role in the fields of biometrics and document forensics. The development in deep learning has led to significant advances in signature verification over the past decade. However, it is still challenging to distinguish between skilled forgeries and genuine signatures because both are close similarities with only subtle differences in strokes. With this paper, we develop a novel squeeze-and-excitation transformer structure (named SET) for feature extraction and signature verification. SET comprises four stages and receives a two-channel signature pair consisting of reference and query signatures as input. The SET block is the core of each stage, which is utilized to enhance the feature learning ability and strengthen the association between feature channels. We evaluate the proposed SET on several public datasets (CEDAR, BHSig-B, and BHSig-H). Experimental results demonstrate that our approach outperforms existing methods.