Combining segmentation by nonlinear diffusion and shape adaptive block-matching for content based coding

A technique for image simplification and segmentation in scale-space is presented. Segmentation masks are then used to estimate accurately the parameters describing motion of single objects in the scene. The segmentation approach relies on a non-linear diffusion model in which multiscale image simplification and subsequent segmentation of the resulting smoothed images is performed. Motion parameters describing the dynamic of each single object are estimated by applying a generalised block-matching approach. The main strategy behind this technique is to use bilinear transformations to establish a spatial correspondence between the points in the input and output images. The performance of the presented techniques is evaluated by processing natural video sequences