Robust text line detection in historical documents: learning and evaluation methods

Text line segmentation is one of the key steps in historical document understanding. It challenging due to variety fonts, contents, writing styles and quality documents that have degraded through years. In this paper, we address limitations currently prevent people from building models with a high generalization capacity. We present study conducted using three state-of-the-art systems Doc-UFCN, dhSegment ARU-Net show it possible build generic trained on wide datasets can correctly segment diverse unseen pages. This paper also highlights importance annotations used during training: Each existing dataset annotated differently. unification its positive impact final text recognition results. end, complete evaluation strategy standard pixel-level metrics, object-level ones introducing goal-oriented metrics.