Text segmentation by integrating hybrid strategy and non-text filtering

The text embedded in images provides important information for image understanding. Text segmentation is an essential step for text recognition. It is often difficult to segment text from images at low resolution or with complex background. In this paper, a novel text segmentation framework is proposed to solve the problem. The proposed framework adopts a hybrid strategy integrating two different text segmentation methods to produce text candidates. One segmentation method is designed based on the intensity uniformity of text regions, while the other is developed by integrating the features of intensity and stroke width of text. To separate text pixels from the text candidates, a new non-text pixel filtering method is proposed. In the filtering method, an effective classifier is designed based on the number of breaking elements and the k-means clustering algorithm. The performance of the proposed segmentation framework is tested by the pixel-based and recognition-based evaluation methods. Experimental results show that the F-score of the proposed framework on the video caption dataset and born-digital dataset of ICDAR2013 are 95.29% and 89.09% respectively, while the correctly recognized character rate and word rate on the German TV public dataset are 91.00% and 72.33%. The experimental results indicate that the proposed text segmentation framework has excellent performance and high robustness in text segmentation and recognition.