Research On Technologies Of Semantic-based Video Retrieval

With the rapid development of science and technology, the amount of video datashows an explosive growth. How to retrieval the video information you want from alarge-scale video database becomes an urgent issue. The current video retrieval systemmainly includes two types: one is based on text retrieval technology; the other is basedon contents video retrieval technology. The former needs manager to annotate videos byhand. However, with the increasing amount of video data, it cost so much human laborsand unrealistic. The latter combines mainly image processing technology and computervision processing to compare the similarity between video clips based on the lowfeatures of image frame, which is very different from that people are accustomed tojudge similarity using the high-level semantic concept. So, bridging the semantic gapbetween low-level visual features and high-level semantic visual concepts and realizingsemantic-based video retrieval may be the biggest challenge that we face in supportingsemantic-based video retrieval.For semantic content-oriented video retrieval, the concept detection is the mostcritical technology. Its purpose is to automatically detect the video that contains a largenumber of basic semantic concepts, such as sky, sunrise, beach, and so on. Through theestablishment of various semantic concept detection models, semantic concept-basedvideo retrieval can be realized. In this paper, a video semantic concept detection methodbased on the evidence theory (D-S) is proposed, Main works are summarized asfollows:For the low-level video features, the three kinds of low level features, color, texture,shape, are extracted from key frames and a support vector machine is trained for eachfeature as classifier. Because the input features and the parameters of SVM will impactthe classification accuracy, genetic algorithm is used for feature subset selection andSVM parameters optimization. Then they are used to model the SVM classifier.Experimental results indicate that the proposed method can obtain optimized featurevector and optimized parameters of SVM classifier, which can improve the performanceand accuracy of single classifier.On fusion algorithms, a video semantic concept detecting method based on theevidence theory (D-S) is proposed, which uses the advantages of combining uncertaininformation and obtains classification results. The SVM classification results based on color, shape and texture feature are regarded as3independent evidences and basicprobability assignment function is constructed through introducing the posterior outputprobability of the SVM. The three evidences of support vector machine are fused by theD-S synthetic rules and the final classification confidence is obtained by the thresholdsof decision rules. The proposed method is compared with maxima method, averagemethod and linear weighting method. Experimental results show that the proposedfusion algorithm improves the precision of concept detection.Finally, based on the semantic concept detection of video key frames, a videoretrieval system based on shots is realized. Through query semantic concept approach,in accordance with the confidences value of relevant shots, from highest to lowest, thevideo shots which belong to specific semantic concept are returned.