Research On Chinese Text Location In Natural Scene Images

Text detection and location in natural images have important significances for the analysis of image’s content, description and understanding of natural scenes. It can be widely used in robot vision, blind navigation and image retrieval and other fields. However, text location tends to be quite challenging due to the uncontrolled text fontsize,color, direction of diverse and complicated background. Text location in natural images remains a challenging technical problems, has been increasingly popular in the field of image processing.In this paper, we focus on Chinese text location in natural scene images based on the current research situation and text location methods. A Chinese text location method suitable for natural scene images is proposed. We combine Maximally Stable Extremal Regions(MSER) with Stroke Width Transform(SWT). We employs MSER as our basic character candidates which include many non-text areas.In the stage of candidate text areas identification and verification preliminary, we define four filtering mechanisms for excluding non-text connected components: restrictions on the height and width of Chinese characters, restrictions on the aspect ratio of Chinese characters, restrictions on the corner of Chinese characters, restrictions on the stroke width of Chinese characters. The experimental results show that they can effectively remove obvious non-text areas after a preliminary filter.After the initial filter, we analysis and integrate the remaining candidate connected component regions. We define three qualifiers: vertical and horizontal interval between two adjacent candidate region is too small, then correspondingly merger them; if the large candidate region contains more than five candidate regions that larger than 32x32, these small areas which may be Chinese characters, we remove large candidate regions, otherwise we remove the small candidate regions inside the large areas; for the overlapping area between two adjacent candidate areas meet certain conditions, then correspondingly merger them.In the stage of determining the text area and non-text area precisly. We use verification method based on SVM(Support Vector Machine, SVM). Due to the characteristics of Chinese characters, we extracted the histogram of oriented gradients(HOG) features, local binary patterns(LBP) features and Gabor features. The combination of HOG features and LBP feature, combination of HOG features and Gabor feature respectively are used as Chinese text categorization feature. And then we verify the text areas based on SVM, Finally, we locate the Chinese text in natural scene images. Experimental results show that the combination of HOG features and Gabor feature classification is better, can be more effectively distinguish text and no-text area.Finally, we collect images by ordinary smartphone terminal and build a natural scene image database that contains 240 Chinese text images, and then we mark them manually. 140 images were used as training samples, and the remaining 100 images were used as test samples, the accuracy of text location is 0.72, the recall of text location is 0.66. The results verify the validity of our method.