Video Text Detection And Localization Based On Statistic Learning

Besides video image information, text is an effective method to accurately express video content. People can understand the main idea only from the video text, which saves mass time spent on watching the video. Moreover, text provides useful information for video retrieval. The video text detection can judge whether text is in the video, which plays a role in screening for text localization and improves the localization velocity and precision rate. The video localization can decrease video data memory capacitance and give a clue for video understanding and retrieval. Therefore, the study on video text detection and localization is very important.Statistic learning theory whose guiding result is the generalization bounds studies on classification, regression and prediction for samples with limited data. The minimax probablity machine and support vector machine et al. is developed based on ststistic learning, which can show the special advantage in solving the machine learning problem with small sample, nonlinear and high-dimensional. The video text detection and localization can be implemented based on statistic learning algorithm. The main work is as follows:(1) Study on video detection based on minmax probability machine. Texture feature information is extracted by discrete cosines transform from video frame as samples, minmax probability machine is trained combining with the judgement of threshold condition and text features, then the distribution of discrete cosines transform blocks is got. Finally, the video frame containing text or not can be distinguished at the place where the distribution of blocks suddenly changes. The experimental result shows that the minmax probability machine is effective on video detection and the precision rate is about 95.2%.(2) Study on video localization based on gradient discrete cosine transform. For locating the video text with simple strokes also studying on the texture feature of text, the gradient operator is introduced in discrete cosine transform, during feature extracting process, which extrusively exhibits text edge information. The candidated text area is dealed with filting and morphologic processing. The experimental results show that the texture feature is aviliable for distinguishing text from background by gradient discrete cosines transform. Compared with discrete cosine transform, this algorithm declines by 2.4% on missing rate for simple stroke and shortens 2.2s on operating time, which can be applied to static or rolling video text localization.(3) Study on video text localization based on fuzzy support vector machine. For accurately locating text in the video with complex background and decreasing the false alarm rate, gray, edge information and texture which is gained as above, is extracted as three-dimensional samples from video frame, fuzzy support vector machine is trained combining with appropriate parameter and kernel function, post-processing is executed by use of video text inherent feature, and exact text region is marked with text box. The membership divides the degree of samples belonging to different classifications, therefore, compared with the classical support vector machine, this algorithm increases by 6.2% on precision rate, reduces by 1.5% on false alarm rate, which can also be applied to static or rolling video text localization.