윤곽 특징과 LAT(Local Affine Transformation)를 이용한 필기체 숫자 인식

Abstract

본 논문에서는 오프 라인 필기체 숫자 패턴에 대하여 LAT(local affine transformation)와 윤곽 특징을 이용한 패턴 정합에 관하여 연구하였다. LAT는 다른 정합 방법들보다 다소 시간이 많이 소요되지만 크기 변화, 회전, 일그러짐, 이동 등의 변형을 흡수할 수 있고, 구조적 특징과의 결합이 용이하다. 가우시안 윈도우 함수의 분포를 줄이면서 LAT를 반복하여 수행함으로써 알고리즘이 수렴하게 되며, 최적 정합을 이루게 된다.

입력 패턴과 마스크 패턴의 윤곽 특징을 비교하여, 유사도가 큰 마스크 패턴들을 선별하여 LAT 알고리즘에 적용시켜서 LAT를 수행하는데 소요되는 시간을 줄였으며, 인식의 신뢰도를 향상시키기 위하여 LAT에 구조적 특징인 교차수와 주축을 LAT의 목?浩獨痔? 가중치로 첨가하였다.

An off-line pattern matching method of handwritten numerals using the local affine transformation(LAT) and structural features is proposed in this paper. Even though the computation time of the LAT is somewhat longer than the other matching methods, the LAT is known as a suitable method to recognize handwritten characters since it is invariant in scale, rotation, shearing and translation, and it is easily coupled to structural features, such as crossing counts, principal angles, etc. The experiment shows that the corresponding mask pattern is converged to its input pattern as the iterations of the LAT are continued.

A preclassification is added with the profile features in order to decrease the computing time. The principal angles and crossing counts are combined to the LAT as the weight of the objective function in order to improve the performances.

The algorithm is applied to the recognition of handwritten numeral images. The average recognition rate is 85.5%.

An off-line pattern matching method of handwritten numerals using the local affine transformation(LAT) and structural features is proposed in this paper. Even though the computation time of the LAT is somewhat longer than the other matching methods, the LAT is known as a suitable method to recognize handwritten characters since it is invariant in scale, rotation, shearing and translation, and it is easily coupled to structural features, such as crossing counts, principal angles, etc. The experiment shows that the corresponding mask pattern is converged to its input pattern as the iterations of the LAT are continued.

A preclassification is added with the profile features in order to decrease the computing time. The principal angles and crossing counts are combined to the LAT as the weight of the objective function in order to improve the performances.

The algorithm is applied to the recognition of handwritten numeral images. The average recognition rate is 85.5%.