High-Level Concept Detection in Video Using a Region Thesaurus

This work presents an approach on high-level semantic feature detection in video sequences. Keyframes are selected to represent the visual content of the shots. Then, low-level feature extraction is performed on the keyframes and a feature vector including color and texture features is formed. A region thesaurus that contains all the high-level features is constructed using a subtractive clustering method where each feature results as the centroid of a cluster. Then, a model vector that contains the distances from each region type is formed and a SVM detector is trained for each semantic concept. The presented approach is also extended using Latent Semantic Analysis as a further step to exploit co-occurrences of the regiontypes. High-level concepts detected are desert, vegetation, mountain, road, sky and snow within TV news bulletins. Experiments were performed with TRECVID 2005 development data.