Video Retrieval System based on Machine Learning

按示例查询(QBE)可以看作是一个机器学习问题。在给定要查询的视频的情况下，构建了一个分类器来根据颜色、边缘和运动等特征区分相关和不相关的视频。本研究从机器学习的角度对QBE进行了探索，包括训练样本、特征、学习算法和数据量等。因此，我们开发了一种快速、准确的方法，可以从大量的视频数据中检索与查询相关的视频。此外，通过将开发的方法应用于对象识别，我们在由NIST赞助的年度全球视频分析比赛TRECVID 2012上取得了最高性能。

Query-By-Example (QBE) can be considered as a machine learning problem. Here, given videos for a query, a classifier is built to discriminate between relevant and irrelevant videos based on features like color, edge and motion. This research has explored QBE from the perspectives of machine learning, such as training examples, features, learning algorithms and data size. As a result, we have developed a fast and accurate method which can retrieve videos relevant to a query from a large amount video data. Furthermore, by applying the developed method to object recognition, we have achieved the highest performance at TRECVID 2012 that is a NIST-sponsored annual worldwide competition on video analysis.