Research Of Video Advertising Text Extraction Algorithm

With the development of network and multimedia technologies, and the popularity of intelligent terminals, recording a variety of video ads and spreading them on the Internet is becoming more and more easier. Therefore video advertising has become the most popular business social media and the most important means to enhance brand awareness. Text in video ads highly summarizes advertising video content, which mainly includes the name of advertising goods, commodities producer, and description of the merchandise function, etc. So extracting text from video ad is helpful for the management of advertisement video data and for assuring the healthy development of the advertising industry through automatically detecting and filtering advertisements sensitive words.In order to attract the attention of the audience, the form of video ads is very complicated in the font, arrangement and color design,etc. Compared with other text extraction technologies, extracting advertising text from video is more challenging.This thesis mainly researches the existing techniques of text location, segmentation and recognition, then analyzes their advantages and disadvantages, finally proposes a method for video ads text location, segmentation and recognition.For text location, this paper mainly researches coarse-to-accurate two-stage method.Firstly, it uses edge features of the image and connectivity domain analysis to locating text area coarsely. Then, the frequency features based on wavelet decomposition and spatial domain features of GLCM are extracted from labeled data sets, which used in algorithm of Co-training to train the support vector machine. Co-training algorithm based labeled data sets solves the issue of noise introduced. For no publicly available video ad data sets, this paper collects advertising text image sets from online video advertising.For text segmentation, this paper mainly studies an automatic seed algorithm, and applies it to iterative graph cut algorithm, finally completes text segmentation.For text recognition, this paper studies and analysis convolution neural network structure in detail, and designs the network structure for text recognition on the basis of text recognition program about convolution neural network and Le Net-5 network model.In addition,in order to recognize Chinese, it makes use of text generator to create animage data set which is used to learn the network structure.