Search

Analysis of handwritten documents is a challenging task in the modern era of document digitization. It requires efficient preprocessing which includes word segmentation and baseline detection. This paper proposes a novel approach toward word segmentation and baseline detection in a handwritten document. It is based on certain structural properties of isothetic covers tightly enclosing the words in a handwritten document. For an appropriate grid size, the isothetic covers successfully segregate the words so that each cover corresponds to a particular word. The grid size is selected by an adaptive technique that classifies the inter-cover distances into two classes in an unsupervised manner. Finally, by using a geometric heuristic with the horizontal chords of these covers, the corresponding baselines are extracted. Owing to its traversal strategy along the word boundaries in a combinatorial manner and usage of limited operations strictly in the integer domain, the method is found to be quite fast, efficient, and robust, as demonstrated by experimental results with datasets of both Bengali and English handwritings.

Analysis of handwritten documents is a challenging task in the modern era of document digitization. It requires efficient preprocessing which includes word segmentation and baseline detection. This paper proposes a novel approach toward word segmentation and baseline detection in a handwritten document. It is based on certain structural properties of isothetic covers tightly enclosing the words in a handwritten document. For an appropriate grid size, the isothetic covers successfully segregate the words so that each cover corresponds to a particular word. The grid size is selected by an adaptive technique that classifies the inter-cover distances into two classes in an unsupervised manner. Finally, by using a geometric heuristic with the horizontal chords of these covers, the corresponding baselines are extracted. Owing to its traversal strategy along the word boundaries in a combinatorial manner and usage of limited operations strictly in the integer domain, the method is found to be quite fast, efficient, and robust, as demonstrated by experimental results with datasets of both Bengali and English handwritings.