Research And Realization Of Image Text Detection

On the Internet, search engines, such as Google, Bing and Baidu have changed history, since they are convenient for everyone's life. But as they are text-based engines, picture and video contents could not been retrieved. With tens of thousands of videos posted on the Internet at each day, the tools of browsing and retrieval of video data become more and more urgent.Image text information extraction system includes detection, localization, tracking, extraction, enhancement and recognition of the text in the image. Generally, it is also divided into two main steps: text detection and text recognition. This paper is mainly about text detection, which includes detection of caption text and scene text.Caption text detection: This paper presents a region-based method. First, an edge operator is used to detect the edge map of input image. Then, the maximum difference range map is computed. Then the map is binarized by a local thresholding method. Through run-length smoothing algorithm the characters are combined to text string. Finally, we locate the text region by using region analysis. The method this paper proposed is clear, fast and robust.Scene text detection: This paper proposed an algorithm based on MSER and SWT. In this paper, a new operator GSWT is used to detect text information in an image. The maximum stable extremal region(MSER) is robust to blur, low contrast and low light, color and texture changes, and GSWT is a reliable method for detection of character stroke. The combination of the two methods can be used to improve the performance of text detection.Considering noise contained in scene images, in this case, the original method will be interfered badly. One side, the more noise is, the more areas of MSER and the more time and space cost by GSWT. Other side, the existence of noise increases the probability of false positive and decreases the precise and recall of text detection. Therefore, we discuss three edge preserving smooth filters: EPSF, guided filter and adaptive manifolds filter. These filters eliminate the noise, and, at the same time, retain the edge information of text. Then the method based on MSER and GSWT is still feasible.In this paper, we have realized two systems that one is a region-based caption text detection system and the other is a scene text detection system based on MSER and GSWT. Each of them achieves high level when compared to previous methods.