| In this dissertation, we explored and developed new techniques and methods for image/video retrieval and classification. Various low- and high-level features have been developed. These include augmented histogram, colorfulness, most prominent color, regions computed from DCT coefficients, focus of attention, motion-activity, motion-magnitude, and cut rate. These features, together with other features such as straight lines, text, and human faces form a powerful set of low- and high-level features for image/video retrieval and classification. Their effectiveness is demonstrated in a content-based image retrieval simulation experiment we performed, a knowledge-based video classification prototype system, and a hyper-linked video retrieval prototype system we implemented. We have also explored the use of straightline features for video sub-classification, and camera transform estimation for basketball games.; The augmented histogram we developed captures information about the “spatial distribution” of pixels, in addition to the intensity or color count. Since the spatial information is computed globally in terms of relative distance between pixels, it is insensitive to image rotation and translation. The knowledge-based prototype system we developed uses a rule-based implementation to classify video into one of five possible classes. The rules capture human knowledge on how to classify video. We present experimental results to demonstrate the effectiveness of this approach. The hyper-linked video retrieval system is based on the concept of human déjà vu. We present a prototype system called DejaVideo, which uses visual similarity to find similar shots. |
