关键词: ：视觉词袋, 加速稳健特征, TF-IDF, 关键帧提取, 开放数据, 检索校准

Abstract: With the rapid development of Internet technology, the number of videos is proliferating in
an exploding way. The traditional text-based retrieval method may bring problems of content missing
and semantic gap, and result in lower retrieval relevance score. Therefore, a video retrieval calibration
method is proposed, which is based on bag of visual words and combines the visual features
extraction technology, TF-IDF technology and the open data technology. First, the HSV-based
clustering algorithm is used to extract the video key frames and the weight vector. Second, speed up
robust features and some other visual features are used to resolve video content missing problem.
Third, the TF-IDF technology is used to measure the weight of keyword, and the open data
technology to obtain the visual features and semantic of the query word, then solve the semantic gap
problem. Finally, our video retrieval calibration algorithm is applied on the Internet Archive data set.
The result shows that compared with traditional text-based video retrieval method, our method has a
15% relative improvement on the average retrieval relevance score.

Key words: bag of visual words, speed up robust features, TF-IDF, key frame extraction, open data;
video retrieval calibration