This paper proposes a method for character string extraction from scene images using direction distribution of edge components and their spatial arrangements. First, we detect edge components using Canny operator as applied to individual RGB channels and label the connected edge components. Second, we count the number of edge points of labeled edge component and calculate the maximum of x-coordinate, y-coordinate, and minimum of them. Third, we remove such edge components that have too few edge points, too large size of width and/or height, and show too large difference between width and height. Fourth, we generate character string candidates by adding up of edge components obtained through individual RGB channels. Fifth, we extract character strings based on spatial arrangements of edge components. Sixth, we extract isolated characters using direction distribution features of edge components. Finally, we extract character strings composed of concatenated characters and isolated characters based on tentative segmentation and character recognition by evaluating character likeness of individual components. Experimental results made on a total of 249 images extracted from ICDAR2003 robust reading and Text Locating dataset “SceneTrialTest” show that the proposed method achieves a recall rate of 71.3%, a precision rate of 64.5%, and an F measure of 67.8%.

109 情報科学部・情報科学研究科 ＞ 学位論文 ＞ 修士論文