紀要論文

K-meansクラスタリングとサポートベクターマシンを用いた情景内カラー文字の2値化と認識

Binarization and Recognition of Color Characters in Scene Images Using K-means Clustering

jpn

18810667

https://doi.org/10.15002/00009584

法政大学大学院情報科学研究科

法政大学大学院紀要. 情報科学研究科編

8

117

122

2013-03

Version of Record

This paper proposes a new method for binalizing and recognizing color characters in scene images. The proposed method consists of the following five steps. (1) A set of candidate binarized images is generated via every dichotomization of K clusters obtained by K-means clustering in the RGB color space. The total number of candidate binarized images equals 2K-2. (2) The optimal number of clusters in K-means application is automatically determined based on the sum of within-cluster variances. (3) We calculate mesh and weighted direction code histogram features from every binarized image and feed them into a support vector machine (SVM) to calculate the degree of "character-likeness." (4) The binarized image with the maximum value of "character-likeness" is output as an optimal binarization result. (5) The optimal binarized image is recognized via the modified projection distance. Experimental results made on the public color character image dataset "the Chars 74K" show that the proposed method achieves correctly binarized character image generation and selection rates of 97.6% and 85.6%, and the final character recognition rate of 63.1%, which are superior to those obtained by the conventional techniques.

Article

AA12222297