Image/Video Text Extraction And Its Application

Text embedded in images and video frames carries important semantic information forimages and video, therefore the technology of text extraction is very important for imageunderstanding and content-based information retrieval systems. Nowadays, many commercialOCR systems make a great success and the technology of text extraction and recognition frombinary image tends to mature. However, most embedded text is surrounded by complexbackground and sometimes accompanied by high noises. These factors have restricted theapplication of OCR and posed great challenges to text extraction from images and videos.To deal with the problems caused by complex background, low resolution and variousstyles of text, we propose a robust split-and-merge text segmentation algorithm in this thesis.To segment detected text precisely and efficiently, the proposed algorithm utilizes not only thecolor information but also the scale information of text strokes. Experimental results show thatthis algorithm can remove most background pixels, and provide a clear binary input image forstandard OCR systems.To remove background efficiently, researchers have proposed lots of approaches. Most ofthem are so complex that text detection and segmentation is quite time comsuming. As a result,many text extraction systems are not practical enough. To overcome the disadvantage, wepropose a video text extraction algorithm based on a time-adaptive color model in this thesis.The proposed algorithm starts up an online machine learning process after simple interactionby a user, and then detects and segments text lines from the video based on the adaptive model.The experimental results show that simple user interactions can improve the performance ofthe text extraction system remarkably. The proposed algorithm is useful for those systems thatrequire very high extraction precision and processing speed, but donot care about introducingsimple user interactions.Besides, as an application of video text extraction algorithm, we develop a system that canremove undesired captions in video. In this system, we first detect and segment text from videoframes, then restore occluded regions through spatial restoration as well as temporalrestoration.