탐색 네트워크를 이용한 무제약 필기체 한글 분할에 관한 연구

Abstract

본 논문에서는 무제약 필기체 한글을 높은 신뢰성으로 인식하는 과정에서 필수적으로 선행되어야 하는 문자분할을 위하여, 기존의 분할방법에서 제안된 적이 없는 문자간 여백공간에서 탐색 네트워크를 구성하여 꺾은선 형태의 다양한 분할경로를 찾을 수 있는 방법을 제안한다. 제안된 방법의 특징은 많은 분할방법들이 사용하는 직선화된 분할경로 보다는 자음과 모음의 조합에 의해 구성되는 한글의 구조적 특성을 고려하여 문자간 여백 공간을 활용하는 꺾은선 형태의 분할경로를 제시함으로써 필기자의 특성에 따른 전처리 과정의 제약을 작게 한다. 전체 분할대상 문자열 중에서 중첩되거나 접촉되어 보다 세밀한 분할이 요구되는 문자열만을 따로 분할대상 블록으로 선택하고, 선택된 블록에 대해서도 적은 수의 특징점을 고려하면서도 다양한 분할경로를 탐색할 수 있도록 네트워크 개념으로 접근하여 탐색해 나가는 새로운 방법을 찾는다. 탐색 네트워크에서 문자간 여백 공간인 원하는 영역에서 분할경로가 탐색되도록 하기 위하여 분할대상 블록의 크기에 따라 예상되는 문자수의 추정하고, 이것에 기반하여 서로 다른 탐색 윈도우가 설정된다. 탐색 네트워크와 탐색 윈도우를 기반으로 가능 분할경로들을 선택하고 이것들 중에서 노드들의 평균 이익, 즉 물동량의 흐름을 최대화하는 개념을 적용하여 최상 분할경로를 찾음으로써 무제약 필기체 한글 분할을 연구한다. 전체 실험 데이터 800개에서 91.4%의 분할 정확도를 얻었고, 분할에 성공한 731개의 문자열 데이터 중에는 개략 분할 단계에서의 수직투영만으로 분할 결과를 얻은 경우가 173개로 21.6% 였다.|A correct segmentation of unconstrained handwritten Hangul is a prior work for word recognition. But, because of difficulty of structural properties of Hangul, most recognition processes ready device for artificial segmentation control in collection process of character image to minimize segmentation problem until present. There were a lot of applications that structural difference of Europe language character is represented in English and Hangul segmentation method that have been studied with do not reflect enough characteristic of Hangul applies as it is segmentation method about character in spite of definition. This became major cause that deteriorate whole performance of system adding ambiguity of recognition unit that is segmented as a result. In this paper, we proposed a new method which is not used by previous method that uses a virtual network in the space between characters for segmentation of unconstrained handwritten Hangul. Characteristic of proposed method is first, characters that segmentation is available by straight line path through vertical projection because width between characters exists in some degree among string inputted at rough segmentation step are used segmentation result. Selected string that more detail segmentation is required being used by segmentation procedure through rough segmentation step, and overlap each other without being so or connect as segmentation target block apart. Second, created search network to receive segmentation path of zigzag line form in empty unfilled space between characters at detail segmentation step. this wished to receive segmentation path of zigzag line form that can get results that human creates segmentation path with hand first of all as process that create of various virtual paths from empty unfilled space between character for character segmentation. Made virtual path created by feature points of each characters connects each other nodes that formed node, and is considered feature point in segmentation target block for search network creation. Also, created search network that considered each node that is feature point too after calculates center of gravity point on inside of each triangles in created network to receive segmentation path that is minute more in created network, and again this existent nodes and virtual path that is minute more by linking each other have. Third, this paper made search attained in area that set search windows and wants. That is, because final segmentation path need not to search segmentation path about all circumferences of search network because should exist between character. Proposed search windows was sets to differ according to literal number that is expected that exist in segmentation target block, and used average character width by standard that expect literal number. In this paper, set search windows of three occasion is a S_(B) < 100 pixel by when is a 100 ? S_(B) < 150 pixel and occasion that is a 150 ≤ S_(B) < 200 pixel divide. Fourth, suggested search procedure that is suitable to considered search network and search windows concept. Gave search condition to pare down calculative burden in process that prevent and searches that segmentation path that do not want at segmentation path searching process through transfer between node for possibility segmentation path search is created. And selected segmentation path to maximize that do sum of the average flow amount expected among possibility segmentation path as best segmentation path. By pure Hangul data set that itself is collected for an experiment presents instance of Hangul name of our country to 40 students and did so that data of total about 800 may be collected, symbol and situation did not become suitable candidate for marriage here and number of string has distribution of a 2~5. In an experiment result, the proposed method could get 91.4% segmentation accuracy from 800 word set including touched and overlapped characters collected from various writers. It is selection process of feature point part considers in priority, childhood that search network is created to be supplemented hereafter. Composition of good search network must make literal structural properties be reflected in this processes because heighten possibility that can search right segmentation path. Method that proposed in this paper should like to be extended to vowel and consonant of Hangul, Should like to be supplemented so that is applicable in English character and number.