映像・音声認識,自然言語処理の適用によるメタデータ生成の作業コスト削減効果に関する考察

We propose a task model that semi-automatically generates scene-based metadata based on mediaanalysis technology such as audio/visual indexing and natural-language processing to reduce the costs of generat-ing metadata.Our task model can shorten the task time by reusing both the results of media analysis and existingtext information such as program scripts.SceneCabinet,a metadata generation and editing system,can automati-cally extract scene-based metadata from videos.The system extracts meaningful video slices and textual informa-tion such as scene titles,synopses,and keywords using natural-language processing based on the results of speechrecognition and video OCR.Moreover,the system can import program scripts and use them to automaticallyextract keywords.SceneCabinet provides an intuitive user operation interface including a video browser with keyimages that are automatically detected based on scene changes,on-screen text,camerawork,speech,and music.Experiments showed that SceneCabinet could significantly reduce metadata generation costs.