Interpretability of an Automatic Handwritten Signature Verification Model
Vinícus Oliveira Barros, Celso A. M. Lopes, Byron Leite Dantas Bezerra · 2023
In handwritten signature verification, convolutional neural networks are used in many configurations, producing vastly different albeit satisfactory results when extracting signature features. One problem posed by the usage of these models is that, depending on the application, it is necessary to point the features that were relevant for the signature classification, which sets a barrier for the usage of these networks. In this work, we propose an approach with Integrated Gradients and Saliency maps with a writer-independent, offline signature verification model to extract regions of relevance from a sample input signature. We also analyze regions where signature features such as initial and final pen strokes, pressure, speed, and connections occur to investigate if a proposed feature-extracting network points to the same criteria that handwritten signature verification experts used. Our experiments show that the initial and final pen strokes are the most commonly-occurring, high-relevance features outputted by the selected attribution algorithms with a relative frequency of 66.6%. These results relate to those obtained by the study of the criteria chosen by experts in a similar setting.