文摘
Character segmentation from text lines in degraded historical document images is challenging due to complex background and non-availability of regular structures of text patterns. This paper proposes a new method based on watershed model for segmenting characters from text lines in degraded historical document images. The proposed method filters out noise pixels by exploring Sobel and Laplacian values of pixels, which results in edges that represent text components. We then propose watershed model for studying non-linear spacing between characters based on the fact that watersheds provide information about water flow and volume of collection of water. Experimental results on different datasets, which include degraded historical document images called Indus documents and other Indian scripts, show that the proposed method segments characters better than the existing character segmentation methods in terms of recall and precision.