Text Line Segmentation on Ancient Egyptian Papyri: Layout Analysis with Object Detection Networks and Connected Components

The automatic localization of text lines is an important step in the analysis of handwritten historical documents. It is a valuable tool for further analysis, such as studying handwriting styles and typefaces or realigning fragments based on a continuation of text lines. This paper tests and compares various architectures, originally designed for object detection tasks, for text line segmentation in ancient Egyptian hieratic papyri. The corpus used in this study presents a significant challenge due to noisy text carriers, complex scripts and layouts, and a highly fragmentary condition. The experiments conducted include tests on transfer learning, data augmentation, and various input resolutions. The results suggest that most object detection networks, particularly the Faster R-CNN and Mask R-CNN architectures, are well-suited to identify lines of text even in very complex or fragmented textual layouts. In addition, we present and evaluate an alternative ‘bottom-up’ approach that utilises binarised documents. This method derives text lines from connected components, allowing for greater control over parameters and improved identification of small ink remnants. Although the numerical evaluation metrics may be lower, this technique offers increased control by the user.