| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 94 === The major purpose of multimedia data retrieval is to hunt the related multimedia data effectively and efficiently. From human perspective, unfortunately, there is usually not enough semantic support to help users get the accurate results by using only low-level visual features as done in traditional studies. Generally speaking annotation can be a solution for enhancing the accuracy of multimedia data retrieval. Video annotation has been considered as a challenging research topic since videos carry complex scenic semantics in addition to visual features. In this paper, we propose a novel method for visual features and frequent patterns exists in the video. Our proposed method consists of two main phases: 1) construction of three kinds of prediction models, namely association, sequential and statistical models from annotated videos, and 2) fusion of these models for annotating unknown videos automatically. The main advantage of the proposed method lies in that both of visual features and semantic patterns are considered simultaneously through fusion approach so as to enhance the accuracy of annotation. Empirical evaluations show that our approach is very promising in enhancing the annotation accuracy in terms of precision and recall.
|
|---|
