Kanji Character Detection from Complex Real Scene Images based on Character Properties

Character recognition in complex real scene images is a very challenging undertaking. The most popular approach is to segment the text area using some extra pre-knowledge, such as "characters are in a signboard'', etc. This approach makes it possible to construct a very time-consuming method, but generality is still a problem. In this paper, we propose a more general method by utilizing only character features. Our algorithm consists of five steps: pre-processing to extract connected components, initial classification using primitive rules, strong classification using AdaBoost, Markov random field (MRF) clustering to combine connected components with similar properties, and post-processing using optical character recognition (OCR) results. The results of experiments using 11 images containing 1691 characters (including characters in bad condition) indicated the effectiveness of the proposed system, namely, that 52.9% of characters were extracted correctly with 625 noise components extracted as characters.