Overlaid Caption Extraction In News Video Based On SVM

With the widespread use of the internet and computer, people get more and more multimedia information everyday. How to index and retrieve digital data effectively promotes the development of the multimedia retrieval and data mining technology. The main information of a story unit in a news video is often carried by the overlaid captions which both supply segmentation clues for video analysis and provide important semantic tags for video indexing and retrieval. Therefore, caption extraction technology plays a significant part in news video retrieval field.A novel overlaid caption extraction approach based on SVM is proposed in this paper. In consideration of the complex background in which captions are embedded and the low resolution of the video during digitalization and lossy compression, a new combination of texture features, which include 20 features in all respectively extracted from different gray level co-occurrence matrices, wavelet coefficients and different oriented edge maps, is presented here to discriminate text and non-text blocks in the shot-cuts. The features are contrast, correlation, entropy, wavelet coefficients variance, wavelet coefficients histogram variance and oriented edge intensity ratio. Then, due to its excellent classification ability, SVM is chosen as the final classifier. The following experiments show that the choices of the features and the classifier are proper and efficient. In addition, considering that the title captions in news video are usually distinguished from those interview ones by colors, a method based on k-means clustering is employed to discriminate the two kinds. At last, post-processing is involved in our approach to prepare for the following OCR, including morphological filtering, projection profile analysis and so on.