Research On Text Extraction Technology In Video

Text presented in videos provides important supplemental information for video indexing technology. In this paper, the key problems in the text extraction in digital video such as the text localization, text tracking and text segmentation are discussed. Because of the complex background, low contrast between the text area and background, the low resolution, the variety in the orientation and the different size of the text, it is challenge to deal with the text in video. Although many efforts have been made in the fields, the results are still not satisfying. This dissertation presents our work on several important aspects of text extraction in video, including text localization, text tracking and text segmentation.In order to deal with these aspects more efficiently, the corresponding algorithms are presented systematically.This dissertation mainly consists of the following parts:(1) A method of text detection based on a wavelet modulus maximum algorithm module and a 2-step classification module is presented. The 2-step classification module is composed of a cascade classifier and a one-class support vector machine (1-SVM) classifier. Most of non-text regions are filtered out by the cascade classifier and the remaining text regions are further verified by 1-SVM. Only the regions accepted both by cascade classifier and 1-SVM are the final text results. It improves the precision of the text localization and reduces the false detection rate.(2) In order to reduce the process time by avoiding doing text localization for each frame in video, and to improve the localization precision, a text tracking method based on robust matching criterion is proposed, and further the multi-resolution algorithm is introduced which reduces the complexitiy in matching process and speeds up the text tracking.(3) In order to deal with the low resolution in the text area in digital video and to deal with the impact of the variety of illumination, an efficient text segmentation method based on sub-pixel algorithm is presented. By introducting the improved Niblack algorithm and connected component based method to the sub-pixel images, a better segmentation result can be got.