Automatic Subtitle Extraction And Recognition In News Video

The caption text in News video frames often contains many important information,such as names of newsmakers, the theme of news events and so on. News Video caption text as a high-level semantic information,they do great help to index News video and understand News video content.This paper studies News video caption text extraction and recognition, including studing the algorithm and key technology about News video caption text localization, tracking, segmentation and recognition.In the step of text localization, this paper uses a text localization algorithm for News video frames in a localization-to-verification framework. In detection and localization step, the algorithm takes full advantage of the information of text edge, through the video frames edge detection and region filter to get the candidate caption text lines. Verification phase applies the method based on region edge similarity which calculates the edge similarity of corresponding caption text lines region between the current frame and its adjacent frame to confirm candidate caption text lines.In the News video caption text tracking and enhancement stage, taking into account that tracking objective is the static caption text of News video, this paper adopts a tracking algorithm based on block matching and edge similarity of text region. For improving the tracking speed, this paper uses the binary search method to track the caption text, so that the tracking process can greatly reduce the cost of localization. In the tracking process, we use the multi-frame fusion to enhance text frame.In News video caption text segmentation stage, this paper applies Otsu to convert enhanced images from tracking step into binarization. Then this paper applies a text segmentation algorithm based on horizontal and vertical projection histogram that can quickly and accurately segment text from the relatively simple background of News video caption frames.Finally, this paper implements a simple recognition system for characters in simplified Chinese primary character set. This system is a three-level recognition system that has coarse classification, fine classification and recognition stage. The system adopts correlation matching method as recognition method. This paper selects mesh feature, peripheral feature and local stroke density feature as classification features.