Análise Comparativa de Modelos para Detecção de Assinaturas em Imagens

Abstract

In real-world applications, such as fraud prevention systems or process automation, the ability to compare signatures between different documents is crucial. The first and most critical step in this pipeline is to accurately locate all signatures present on a scanned page. This paper addresses this initial stage by conducting a comparative analysis of six modern object detection models: YOLOv12, DINO, RetinaNet, Faster R-CNN, Double Heads, and VFNet. Using a public dataset of annotated documents and a standardized experimental protocol, we evaluate which model is most effective for the specific task of signature detection. The results provide a solid foundation for developing robust automatic signature analysis systems, indicating the most suitable architectures for this essential detection phase.