Research On Segmentation Algorithm For Handwritten Nǚshu Characters Image

Nǚshu is a unique character set which is only used for women in the world. Nǚshucharacters are passed from generation to generation by the handwritten way. With the Nǚshuinheritors died successively, Nǚshu is in an extremely difficult situation and urgently needs tobe protected by using information technology. The research of off-line Nǚshu charactersrecognition is of great significance to rescue and protect Nǚshu. Nǚshu characterssegmentation is the foundation of off-line handwritten Nǚshu characters recognition. Theaccuracy of segmentation affects the rate of recognition directly. The research of Nǚshucharacters segmentation has important meaning for improving the overall system performanceof off-line handwritten Nǚshu characters recognition.There are many carriers for Nǚshu, and the writing style is special. The text lines inNǚshu images have multi-oriented structure which is difficult to extract. Nǚshu is writtenfrom top to bottom and there are up-down overlapping and up-down touching characters inNǚshu images. These issuses bring great difficulties to the segmentation of Nǚshucharacters, and affect the following information processing indirectly. According tothe characteristics of Nǚshu characters, the approach for extracting multi-oriented textlines and the method of character segmentation for Nǚshu are designed and realized inthis paper. The major researching contents in this paper are as follows:(1)The multi-oriented structure such as diverging or bending structure is widelyexisted in the text lines of off-line handwritten Nǚshu character images. To overcomethe problem, an extracting method for multi-oriented text lines based on the linkmodel is designed and realized in this thesis. Firstly, morphological operation and minimumenclosing rectangle are used to extract blocks of Nǚshu characters and delete non-characterblocks. Secondly, a triangulation network is built by Delaunay triangulation on these blocksof characters. According to three given rules, a link model is constructed by calculating theweight of each edge in the triangulation network. Finally, according to the weight of the linkmodel, the optimal text lines are extracted by mutual exclusion. The experimental resultsshow that our method can effectively extract the multi-oriented text lines from Nǚshucharacters images on different carriers, such as fans and handkerchiefs.(2)To overcome the problem of up-down overlapping and touching in the text imageabout Nǚshu characters, a multi-step segmentation method based on thinning method isdesigned and realized in the thesis. Firstly, the method employs histogram projection to complete the pre-segmentation. According to the average height of characters, overlappingand touching characters are selected. Secondly, a path searching method based on backgroundthinning is proposed in this thesis which can solve the overlapping characters. Finally, thethesis presents a detection method of adhesion points which turns touching Nǚshu charactersinto overlapping characters, so that it can use the overlapping method to cut touchingcharacters. Experimental results show that the proposed segmentation method caneffectively segment the up-down overlapping and touching Nǚshu characters.