| The rapid development of information technology promotes the rapid growthof video data in the network, the news video which has relationshipswith people basic necessities of life as one of the important video has been widelyconcerned. However, with the rapid growth of news video data,how tomanage, organization and retrieval of news video effectively has become the urgentneed to solve the problem. The traditional news video retrieval methods relyon manual annotation and underlying feature, however, people also hope achieve theretrieval of news video through the semantic level. therefore, semantic analysis ofnews video are the basis for effective retrieval of news video. This article mainlyanalysis the semantic of news video from three aspects which is structure of newsvideo semantic, theme semantic and high-level semantic. the main work of thispaper are as follows:(1) This paper presents a detection method of the boundary of news videobased on color weighed, the news video is a series of video frames, so the newsvideo shot segmentation can simplify the processing work to the video, the accuracyof shot boundary detection directly affects news video semantic analysis. Thisalgorithm calculated the difference between frames feature used theHSV histogram weighted. And got the news video shot boundary compared withthe adaptive threshold, because the flash phenomenon occurs frequently in newsvideo,so this paper proposed a improved sliding window method to filter the effectof flash. and the experimental results show that the recall andprecision the boundary of news video has a high detection. In addition,basedon news video boundary detection, according to the editing features ofnews video,the paper completed detection of anchorperson shots based on facedetection.(2) News video visual features reflects the key information to some extent. Inthis paper, the main color is selected as the main color feature. Texture featureis gray level co-occurrence matrix, and the energy, entropy, moment of inertia andthe correlation. This paper presents the color and texture features are nonlinear fusion, greatly reduced the dimension of the visual feature vectors. And useSVM (Support Vector Machine) classifier to classify the news video. Throughthe experiment. Visual features and textual features of nonlinear combination canimprove the accuracy.(3) This paper proposes a manner of using horizontal scanning and verticalscanning to detect the caption region, looking for continuous multiple lines ofadjacent pixel gray difference square sum is greater than the threshold regionas candidate caption region in the horizontal scan, and using a vertical scan cuttingsubtitle ends. And completed the caption area preprocessing, character segmentationand feature extraction, finally detect the text recognition using neural networktheory.(4) Finally, the paper get the final feature vector by the vector text and visualvector linear combination and detecting the corresponding keyword to determine thetext vector. according to the SVM (Support Vector Machine) theory, the method ofdouble modal semantic classification of news video obtain good experimental effect.To sum up, this paper systematically studies the research on newsvideo semantic analysis, after the semantic of structure and semantic of theme,studied the high-level semantic of news video by combining text andvisual characteristics. |
PDF Full Text Request