Determining the Segmentation Type Impact on an ID Card Fraud Detection System

Fraud detection on ID card remote presentation helps determine if a government-issued ID card has been doctored before enabling the user to perform banking and e-commerce activities from home. In the literature, using a semantic segmentation stage before the Fraud Detection Network (FDN) is a common practice. In this way the FDN focuses only on the contents of the ID card to make a prediction. However, this work aims to study whether a different kind of segmentation could allow the FDN using more contextual information of the background to improve performance. We separately trained and tested two FDN architectures with four types of cropping and segmentation. Our models obtained BPCER \(_{\textrm{100}}\) scores of 5.40% and 3.12% respectively. The best performing network achieved that score training with cropped images, which contain a small portion of the background. This indicates that contextual information can increase performance and reduces processing time by using detection networks instead of semantic segmentation.