Semantic concept detection from visual content with statistical learning

Abstract

This thesis addresses semantic concept detection from visual content with statistical learning methods. The highest level semantic concept is genre, the lowest one is object. Accordingly, two parts of research work have been conducted, namely sports news video genre identification and automatic image annotation. For the former, two novel features were proposed to classify sports news video shots; due to the variation of content and shot length, this problem is high challenging. For the latter, we proposed a novel automatic image annotation framework and achieved promising results which outperform the state of art works in two famous dataset: CorelCD and TRECVID2003. Our contributions can be summarized from two aspects: first, proposed a novel image representation scheme with which an image can be treated as a text document, so many text document techniques can be employed; second, proposed two flexible information fusion methods for fusing diverse visual features and multiple modalities.