Research On Segmentation Of Unconstrained Handwritten Characters

Character segmentation is a critical sector in the total recognition system since anyerror in this stage will propagate to all later analysis. Resent researches and applicationsshow that character segmentation has become the core to centralize character classification,recognition kernel and context postprocessing. And study on how to saperate characterscorrectly before they are sent to the recognition engine is very significant to improve thetotal system performance.In this dissertation, several key techniques and algorithms on handwritten characterssegmentation are discussed.Handwritten Chinese characters may be written touching or overlapping each other,and any single method can not give perfect solution to them concurrently. For abovesituations, a multi-stage approach is proposed for off-line handwritten Chinese stringsegmentation, which consists of non-touching characters segmentation, touching characterssegmentation and over-segmented characters merging. By means of combining severalmethods effectively, each problem is solved in a certain stage and total performance isimproved.In the procedure of merging over-segmented characters, the well-known A~* heuristicsearch algorithm is applied to find the optimal segmentation results in the weightedsearching graph. With dynamically determining the the minimum segmentation costfunction as the criterion of cortfidenee level, the number of the traversed paths would bemininal.Drop fall algorithm is more accurate in several contour splitting algorithms, but it islacking in locating the start point in sliding process and seeping process in verticaldirection through the black block. For the purpose of solving the problems and gettingbetter performance, background region analysis is proposed for segmenting handwrittentouching numerals. According to the information extracted from the given backgroundregion, the location of the start point is specified, and the criteria are established todiscriminate the type of the touching numerals. To split the numerals, different strategyoriented touching type is implemented by selecting an appropriate algorithm from the setof Drop fall algorithm, and the seeping process is extended and improved to get accuratesegmentation results, when the drip moves through the junction between the two numerals.The problem of segmenting the cursive handwritten letters is made complex by the factthat the writing is inherently ambiguous as the letters in a string are generally linked together, poorly written, and may even be missing. As a consequence, cursive letterssegmentation requires sophisticated techniques. Therefore, an approach based onrecognition and postprocessing is presented, which adopts the prevalentovers-segmentation technique and makes full use of recognition and contextualinformation to compensate for the ambiguity. Firstly, the string image is divided into somesegmentation regions, each of which contains a splitting path to assure no more than onecharacter in it. In the following step of generating the splitting path by dynamicprogramming, the information of gray scale and binary images are combined in order toavoid the limitation of both gray-scale image and binary image. In the recognition stage, theover-segmentation verifier is designed for determining whether each presegment is furtherprocessed or not, and the verifier interacts with the classifier and the statistical languagemodel to achieve more reliable results, namely each segment containing exactly one letter.