High level semantic concept retrieval using a hybrid similarity method

In video search and retrieval, user’s need is expressed in terms of query. Early video retrieval systems usually matched video clips with such low-level features as color, shape, texture, and motion. In spite of the fact that retrieval is done accurately and automatically with such low-level features, the semantic meaning of the query cannot be expressed in this way. Moreover, the limitation of retrieval using desirable concept detectors is providing annotations for each concept. However, providing annotation for every concept in real world is very challenging and time consuming, and it is not possible to provide annotation for every concept in the real world. In this study, in order to improve the effectiveness of the retrieval, a method for similarity computation is proposed and experimented for mapping concepts whose annotations are not available onto the annotated and known concepts. The TRECVID 2005 data set is used to evaluate the effectiveness of the concept-based video retrieval model by applying the proposed similarity method. Results are also compared with previous similarity measures used in the same domain. The proposed similarity measure approach outperforms other methods with the Mean Average Precision (MAP) of 26.84% in concept retrieval.

Actions (login required)

View Item