Research On TV Video Caption Recognition And Retrieval Technique

In the information age today,the growth of the video data content is explosive,which makes the automatically collation of massive video information become an urgent need in the contemporary academia and industry field.Video captions have a strong correlation with the video content,and carry rich high level semantic information.By studying the video caption detection and recognition issues,this paper establishes a video retrieval system.Our work is as follows:Firstly,we propose a video caption binarization algorithm based on high contrast images.After analyzing the common features of video caption text,this paper uses an adaptive local contrast algorithm to obtain high contrast image,and binarizes it based on OTSU method and the statistical distribution of pixels' grayscale.Secondly,we focus on the character segmentation method.In order to get the accurate positioning and segmentation of characters,this paper analyzes the characteristics of Chinese characters and common segmentation errors,and then segments the characters using word-wide clustering.According to the features that video caption remains for some time in the frame stream,we de-noise the detection text by fusing images between frames.Third,in order to retrieve a large amount of video data quickly,this paper proposes a structure analysis of video data centered on captions,and extract corresponding key frames using video shot detection.And the introduction of inverted index and space vector model enables the efficiency of retrieval system be greatly improved.Forth,this paper presents a front and end architecture of video caption identification and retrieval system.Front-end system is responsible for filtering the video stream for text extraction and recognition,achieved by a PC or a DSP,the recognition result backhauls to the backend server for indexing and other operations.Experimental results show that the proposed method have good results for a variety of styles caption text.This paper establishes a video test data set,and test results show that with an 84 percent video caption identification accuracy,the system still has a good real-time performance,and has a good potential of multi-channel parallel processing.