Research On News Video Semantic Concept Detection

With the rapid development of network and communication techniques, the amount of news video available to ordinary users has significantly increased, and effective retrieval of massive news video is urgently needed. Different from the high-level conceptual interpretations of people, traditional content-based news video retrieval methods use low-level feautures to represent the content of news video, while people are more willing to retrieval news video at semantic level. News video semantic concept detection is introduced to bridge the semantic gap between low-level features of news video and the high-level conceptual interpretations of people, which is the foundation of semantic-based news video retrieval. In this thesis, news video semantic concept detection techniques are studied in depth, and three main contributions are listed as follows:(1) To resolve the problems of traditional BoVW(Bag of Visual Words) method, such as visual word synonymy and polysemy and the time efficiency decreasing as the visual data scales up, a news video semantic concept detection method based on randomized visual dictionaries is proposed. Firstly, E~2LSH(Exact Euclidean Locality Sensitive Hashing) is used to cluster local feature points of training video key frames, and a group of scalable randomized visual dictionaries are constructed. Secondly, for each semantic concept, a group of SVM(Support Vector Machine) classifiers are trained based upon these randomized visual dictionaries. Finally, a voting strategy is employed to integrate the opinion of each SVM classifier, thus accomplish video semantic concept detection. Experimental results show that, compared with traditional BoVW method, the novel method achieves higher semantic concept detection accuracy, while guaranteeing acceptable time efficiency.(2) To make full use of both global and local visual features of news video, an evidence fusion based news video semantic concept detection method is put forward. Firstly, features including grid color moment, wavelet texture and visual word histogram are extracted from key frames of news video, and single feature based SVM models are trained. Then, the generalization error of each SVM model is analyzed, and evidences are generated using a discounting coefficient method. Finally, these evidences are fused with an evidence fusion equation, and the fused result is regarded as the final semantic concept detection result. Experimental results show that the new method improves the detection accuracy and outperforms the traditional linear classifier fusion based method.(3) According to the semantic concept detection algorithms presented above, concept detectors are trained, and a practical semantic-based news video retrieval system is developed using these concept detectors. The experiment conducted in practical environment shows that the semantic-based news video retrieval method can achieve higher recall and precision compared with traditional low-level features similarity matching based news video retrieval methods.