engleski

Scene text extraction using modified cylindrical distance

Text extraction is important step that strongly influences on final recognition result. In case of scene text this task is especially challenging because of variable illumination, complex backgrounds, blur, viewing angle etc. In this paper we propose new method for scene text extraction in HSI color space using modified cylindrical distance as homogeneity criterion in region growing algorithm. Solution for seed pixel selection based on horizontal projection is also introduced. Results are presented for 3 homogeneity criterions: modified cylindrical distance, cylindrical distance and Euclidean distance. It is also shown that combination of these distances improves text extraction performance

scene text; region growing; cylindrical distance

nije evidentirano