Linear and nonlinear scale-spaces for video indexing and retrieval

An extension of the conventional linear and nonlinear scale-space models for shape and image simplification, indexing and retrieval is presented. The linear model can be used for shape based retrieval when the main video objects have been identified. In this context an algorithm for contour simplification is introduced. Two different nonlinear filtering techniques for image simplification are also described. The first one is generated by convolution with a group of anisotropic weighted filter kernels. The second is based on a parabolic differential equation in divergence form. The linear and nonlinear form of these diffusion filters allows to integrate additional information to control the evolution in both spatial and temporal directions. This property is used to extend the conventional nonlinear scale-space in order to perform content-based segmentation of natural video.