The Impact of Hand‐Crafted and Data‐Driven Feature Extraction on Signature Identification: A Comparative Study

Signature identification is the process of verifying the authenticity of a signature by analyzing its unique characteristics. In this paper, we have done a comparative study to investigate how hand‐crafted and data‐driven‐based features could impact signature identification. In hand‐crafted features, we extract the key points using a well‐known scale‐invariant feature transform (SIFT) algorithm. Then, we create two different representations for the signature image: (1) a graph representation, which shows the structural relationship between keypoints; and (2) a visual word histogram representation, which ignores the structural knowledge. In data‐driven features, we fed a raw image into two pre‐trained residual network (ResNet) and VGGNet to extract local structural features from the images. Then, models were fine‐tuned by our data. Two well‐known datasets, the MCYT‐75 and the UT‐Sig, datasets, are used to assess the approach. The obtained results show that the graph representation for the sign image has a more discriminative ability to differentiate the original signatures from the forgeries.In this paper, we have done a comparative study investigating how hand‐crafted and data‐driven‐based features could impact signature identification. In hand‐crafted features, we extract the key points using a well‐known SIFT algorithm.