A Modified U-Net Approach for Word and Line Segmentation in Kannada Palm-Leaf Manuscripts

Palm-leaf manuscripts represent an ancient cultural and literary heritage, however, are difficult to be digitized because of uneven textures, degradation of the leaves, and fine handwritten text. This study provides a consistent text segmentation model based on an enhanced U-Net architecture in order to isolate green-ink text in high-resolution manuscript photographs. The pipeline involves post-processing in order to isolate clean text fragments, ground-truth mask construction by means of automation, thick mask refining, and preprocessing (grayscale conversion, CLAHE enhancement, and binarization). The proposed U-Net model performed well with a Dice of 0.878 and IoU of 0.784 on the validation set after it was trained and evaluated on 1352 hand selected image-mask pairs. Comparing it to the YOLO object detection framework shows that U-Net is always able to produce pixel-precise masks, whereas YOLO fails at this task of segmentation and cannot detect thin and jagged text lines. The results of the experiment indicate that the proposed UNet pipeline is a reliable automated solution to the extraction of manuscript text, which opens the way to improved OCR and digital preservation procedures.