Research Of The Text Extraction And Removal In Video With Complicated Background

With the widespread use of computer and multimedia technology, information of digital video and image has become the mainstream media of information exchange, which has greatly influenced the lifestyle of people. The captions in video images contain rich information of semantics. It is very helpful for the understanding and analysis of video image if we can detect and locate these information. At the same time, it has great significance on video reuse if we can effectively remove the captions embedded in images and restore the covered background.The study of this paper mainly focuses on the detecting and repairing of the text in video image with complex background. My main work contains:In the research on the text extraction, considering the color of text and edge features, this paper propose a method of text extraction based on the color edge detector and connected domain features analysis, which effectively reduce the false detection rate when the gray scale of text is close to background. Meanwhile, the proposed algorithm has stronger robustness and adaptability for text of different types.In terms of image repairing, this paper introduces the classical Criminisi's repair algorithm based on texture synthesis, and then analyzes the structure of this algorithm and its existed insufficiencies. On this basis, the three improvements are put forward.First, the size of the template block:according to the distribution of the image texture, this paper use different sizes of template. Texture-rich region have a smaller template, and texture weak area have a larger template, effectively ensuring the image texture information while improving the repair rate.Second, modify the repair sequence:the multiplication operation of the priority formula is changed into the addition operation, and the weight of data items, that is the structure information, are increased, which effectively reduce the sensitivity of the multiplication factor for zero while getting better repair effect for image of strong structure. Compared with the texture information, the human eye is more sensitive to structural information, so the improved image is more consistent with human visual effects.Third, refine the best match function:considering the local similarity criterion, this paper introduces the concept of distance in the calculation of matching function, which effectively reduce the wrong matching operation and ensure the accuracy of repair.This paper carries on a large number of comparative experiments about the improved algorithm and classical Criminisi algorithm. Experimental results show that the proposed algorithm in this paper have high repair efficiency while maintaining the image texture and structure information.