Semantic Concept Detection for Multilabel Unbalanced Dataset Using Global Features

Digital evolution in capturing video, advances in compression technology and internet leads to availability of large videos on the web. There is growing need for efficiently retrieving relevant videos. Semantic Concept detection assigns multiple labels to segmented shots or entire video which facilitates many applications like multimedia indexing and retrieval. This paper presents the semantic concept detector architecture for unbalanced dataset which assigns multiple labels with probability to input video. The proposed architecture uses visual features extracted on global scale. The unbalanced dataset problem is handled by partitioning dataset into segments further evaluating classifiers on these dataset. Feature fusion and decision fusion is evaluated using machine learning algorithm for all segments. Performance of the concept detection architecture for above fusion methods are reported with Mean Average Precision. The proposed method for multilabel concept detection is evaluated on TRECVID 2007 dataset and performance is better than existing early and late fusion.