Abstract

Content-based image retrieval systems often rely on visual image features such as?color, texture, shape, etc. to extract images ~Ssimilar ~S to a query image from the image?database. Such features are also used for video shot detection and key frame extraction?for searching video databases. After a brief review of some of the existing techniques,?this presentation will focus on an image retrieval and video shot detection scheme?developed at I. I. T. Kharagpur. We have used HSV color histogram to develop a?content based image retrieval system. The HSV representation scheme has the?advantage that it separates luminance and chromatic components. Moreover, by taking?note of the physiological properties of cones and rods in the eye, a 2-D HSV histogram?that captures gray and true color components of the image is constructed for supporting?color based image retrieval. With a view to further improve the performance, a?composite color and texture feature vector (coltex) has also been examined. The HSV?gray and true color features are found to be quite useful for shot detection and key frame?extraction in video retrieval applications. The talk will also cover a video data modeling?scheme developed at I. I. T. Kharagpur, that adopts object-oriented approach for?representing the structure and behavior of the entities of interest. The objects in the?proposed model can be static or dynamic. In order to capture the dynamic behavior of an?object, the state chart based formalism has been adopted. The proposed scheme?enables us to capture dynamic behavior of complex processes (both periodic and?aperiodic). Since the video indexes are dictated by the state transition behavior of the?underlying objects, content based retrieval of video segments based on the time varying?features of the objects can be carried out. A logic-based framework has been developed?to capture the semantics of the video segments. Based on this work we have developed?a video data management and analysis system (VIMALA) that supports UML like user?interface to capture relationship among the video objects. It has been applied to a?number of application domains ranging from echocardiogram videos to sport videos.