Text Line Segmentation on Myanmar Handwritten Document using Average Linkage Clustering Algorithm

Text line segmentation from document images is a significant challenge in the field of document image analysis. It involves extracting individual text lines from Myanmar handwritten document images to enable text recognition. This task becomes particularly challenging in Myanmar handwritten documents, especially those with irregular or cursive writing styles, due to variations in line spacing, and touching and overlapping characters in Myanmar handwritten documents. This paper proposes a text line extraction method based on an average linkage clustering algorithm for handwritten document images to address segmentation errors caused by characters with inconsistent spacing, different writing styles, and line overlaps due to ascenders and descenders. In this paper, Connected Components (CCs) are extracted by using Connected Component Analysis (CCA) and Anisotropic Gaussian multiscale technique. And then convex-hull computation based on the divide and conquer method is used to re-segment the irregular touching components. Then the text lines are extracted by the proposed system based on an average linkage clustering algorithm that consider both the smaller and larger within-cluster variance. The performance of the proposed method is evaluated using the Pixel and Line Intersection over Union (IU) values, which are found to be 93.27% of Pixel IU and 95.09% of Line IU on dataset 1 and 92.61% of Pixel IU and 89.90% of Line IU on dataset II, respectively. According to the experimental results based on the existing dataset and their own data set, the proposed system can give a better result than the Density-Based Spatial Clustering and Application with Noise (DBSCAN) clustering algorithm.