Rolling And Non-Rolling News Subtitle Location And Segmentation

In the times of multimedia, news video is one of main way to obtain information for human. With the increasing amount of information, the video classification and retrieval is becoming increasingly important. In recent years, more and more rolling subtitle was applied to the news video. How to deal with the location, segmentation and recognition of news video with rolling subtitle has became new research direction and topic.Generally, the recognition of news video with rolling subtitle is divided into six steps: detection and location of rolling subtitle, segmentation of rolling subtitle, subtitle frame detection, location of subtitle region, subtitle segmentation, identification of subtitles.In the news subtitles detection, we firstly detect and locate the rolling subtitle area. Based on the features of the rolling subtitle region, we detect and locate rolling subtitle region by calculating the vertical edge of the image and finding edge line. Considering the changing of rolling subtitle, we segment the rolling subtitle by setting a gate and calculating the vertical projection. In this way, we solve the segmentation of rolling subtitle effectively.For the detection of subtitle frame, we detect the subtitle frame by calculating local brightness difference of two adjacent video frames.For the location of subtitle region, considering the edge, brightness, arranged and connected domain characteristics of news subtitles, we locate the subtitle region by extracting image edge and analyzing the brightness, horizontal projection and connected domain of the image.For the binarization of subtitle, we binarize rolling subtitle region by Otus global threshold. Based the feature of non-rolling subtitle, we use a local Ostu threshold method in non-rolling subtitle region, and effectively improve the binarization results.For the segmentation of subtitle, we split horizontal character by vertical projection of characters. For the problems of projection segmentation, we solve the adhesion of Chinese characters problem by comparing word width and we effectively improve segmentation accuracy of the English, Chinese and number mixed subtitle by multi-feature and multi-level model.For the subtitle character recognition, we use a PCA-based subspace learning method. By this method, we improve subtitle character recognition rate and overcome the binarization difficulties caused by poor image quality in the traditional OCR recognition.Experiments show that ours methods work well in the rolling subtitle detection, location and segmentation, the subtitle frame detection, subtitle region location, segmentation and subtitle characters recognition, and has some practical value.