Introduction Previous studies on recognition of handwritten numerals of other languages revealed that about 15% of such characters involved connected numerals [1]. For recognition of connected characters, an efficient fragment technique is essential, while the conventional approaches attempted to separate the string into individual characters by splitting the string slantly and apply distinguishing analysis to retrieve the characters by shapes [1]. Some researchers considered ways to separate

more »

... e characters completely by selecting favorable plausible areas to the segmentation from the background [2]. Some others tried to separate the lexicon string by a continuous density hidden Markov model and used dynamic programming approach to match the character with word images and strings [3]. These approaches might not be able to achieve a satisfactory result. While the uncertainty occurs due to the lack of knowledge on a string with connected characters, the previous algorithms had to make the separation of touched characters by applying an previously assigned confidence value for the segments of words to reflect the ambiguity among character classes, and such separation algorithm could only successfully separate the strings with two-touched or three-touched together characters [4]. In consideration of human-beings ability to deal with the recognition of strings with connected characters, the mixed process of bottom-up (image content based analysis) and top-down (knowledge based recognition) and the combination of segmentation and recognition processes had been proposed by Chi [3], [5] and other scientists for a more efficient algorithm. In their proposed solutions, segmentation boundaries in the forms of vertical or slant lines and in many cases curves [9] or piecewise lines are in utilization for separation of the connected characters. Many researchers have also proposed models of fuzzy logic rules [6], neural networks, linear discriminant functions, template matching, and binary comparisons for evaluation and retrieval of features from the strings with single and double-touched handwritten numerals [7]. Some other good models have been proposed in this area. T.M Ha et al. [14] suggested a system for off-line recognition of handwritten numeral strings, which can recognize a previously divided string fragment of numbers by calculating of weighted sum of confidences over all classes. A very effective algorithm was proposed by Shi and Govindaraju [15], that segmenting connected handwritten digit strings. They have very good results to recognize samples