Research On The Svm-based Video Semantic Extraction And Relevance Feedback

To narrow the"semantic gap"is a great difficulty in video semantic retrieval. High-levelsemantic extraction and Relevance Feedback (RF) are currently promising methods for resolvingthe"semantic gap"problem. SVM has excellent classification performance for finite sizetraining samples, which could be considerably suitable for video retrieval with small size samplefeature, and has been greatly applied in multi-media field. In this paper, our research andanalysis will focus on the applications of SVM in video semantic extraction and relevancefeedback.After summarizing the current research on video high-level semantic extract approaches andanalyzing the successful application of mutli-variable support vector machine regression(MSVR)in the multi-input and multi-output (MIMO) systems of engineering field, we propose to utilizeMSVR method for automatic semantic annotation, then construct semantic featuresautomatically based on those intelligent annotation semantic set. We also compared our MSVRannotation approach with other two methods, SID and Gaussian Kernel algorithm, and fromexperiment can draw a conclusion that our method has slightly better performance than the othertwo. MSVR can implement valid semantic annotation in the case of adequate samples.After analyzing the application of SVM in RF technique, we utilized SVM-based RF toestablish a key-frame semantic retrieval system on the basis of semantic features that constructedabove. Then we improved the conventional SVM-based RF method aims to overcome itsdeficiencies. On one hand, for the small sample size limitation, we proposed a method thatcombined accumulating samples with expanding the association sets, thus to increase thesamples. At the meanwhile memorize and accumulate sample sets to reduce user's evaluationmark fatigue, and optimize the negative sample selection to obtain a trade off between positiveand negative samples. On the other hand, as for the limitation of can't long-term retain the user'sfeedback information, we separately store the result set and the SVM model when achieving asatisfying results, thus to generate historic query record and make the current feedback information be shared. The system can expand association sets of the relevant samples and useSVM model to predict a fast result by making use of historic query records. Thus establish along-term memory function for SVM-based RF.Finally we build a video key-frame semantic retrieval system with long-term memoryfunction. The system can implement automatic semantic annotation, semantic feedback retrievaland content-based retrieval. We use datasets of the standard benchmark TREDVID, and designtwo groups of experiments, the content-based retrieval with non-improved semantic retrieval,and the improved semantic retrieval with that non-improved one, to compare and analyze theirperformances. Additionally, we observe their retrieval results based on different low-levelfeatures, i.e. individual color feature, individual texture feature and integrated feature. Theobservation verified that retrieval with combined feature can obtain slightly better performance.It also indicates that, the precision of semantic retrieval are considerably higher than that ofcontent-based retrieval, and our improved SVM-based semantic retrieval has enhanced itsperformance in both query precision and retrieval effiecincy compared to the non-improved one.