자연 영상에서의 문자 패턴 추출

Abstract

본 논문에서는 자연 영상에서 문자의 일반적인 특징을 이용하여 문자 패턴을 추출하는 방법을 구현하였다.

문자의 크기, 형태, 위치, 농도, 문자 종류, 서식, 그리고 조명의 조건이 일정하지 않은 자연 영상을 대상으로 다음의 단계를 거쳤다. 먼저, 조명의 조건에 크게 영향을 받지 않는 국소적 2치화 방법을 이용하여 배경 부분과 문자 영역의 영역 분할을 행하였으며, 문자와 인접해 있는 배경 부분은 농도 차가 크다는 성질을 이용하여 비문자 영역으로 제거하였다. 또한, 입력 영상에서 문자의 기본적인 성질, 즉 위치, 크기, 폭, 종횡비 등을 만족치 못하는 비문자 영역을 제거하였으며, 문자 내의 구멍의 수에 기초하여 비문자 영역을 제거함으로써 문자 패턴 후보를 추출하였다. 다음으로 문자를 둘러싸고 있는 테두리 부분을 제외한 문자 패턴 후보에 대하여 투영 방법을 적용하여 문자열을 검출하고, 이 문자열에서 각각의 문자 패턴을 추출하였다.

여러 가지 서로 다른 상?꼭? 입력 영상을 대상으로 이 방법을 적용한 겨로가 대부분의 문자 영역 추출이 가능하였으며, 비문자 영역을 효과적으로 제거할 수 있었다.

In this paper, an effective algorithm for the character pattern detection using fundamental features of natural scene images is proposed . The character sizes, forms, positions, intensities, kinds of characters, text styles, and illuminations are not uniform in natural scene image. An image is segmented into background regions and character regions by the adaptive local thresholding method, which is independent of the illumination. The backgrounds of input images are roughly removed by using the characteristic that the gray levels of adjacent regions around characters and those of backgrounds are very different.

For the finer removal of backgrounds and non-character regions, the basic characteristics of character regions are used, that is , the position, size and the aspect ratios of circumscribing rectangles are almost uniform in character regions, and the number of holes of that regions are small, are applied. After the removal of backgrounds, the projection profile method is used to detect character strings. Each character pattern is then detected form the character strings.

The experiments are carried out for several natural scene images, and character patterns are detected clearly with the proposed algorithm.

In this paper, an effective algorithm for the character pattern detection using fundamental features of natural scene images is proposed . The character sizes, forms, positions, intensities, kinds of characters, text styles, and illuminations are not uniform in natural scene image. An image is segmented into background regions and character regions by the adaptive local thresholding method, which is independent of the illumination. The backgrounds of input images are roughly removed by using the characteristic that the gray levels of adjacent regions around characters and those of backgrounds are very different.

For the finer removal of backgrounds and non-character regions, the basic characteristics of character regions are used, that is , the position, size and the aspect ratios of circumscribing rectangles are almost uniform in character regions, and the number of holes of that regions are small, are applied. After the removal of backgrounds, the projection profile method is used to detect character strings. Each character pattern is then detected form the character strings.

The experiments are carried out for several natural scene images, and character patterns are detected clearly with the proposed algorithm.