Image morphing using scattered feature interpolations

Abstract

Image morphing deals with the metamorphosis of an image to another image. Given two images, the metamorphosis generates a sequence of inbetween images in which one image gradually changes into the other over time. Image morphing techniques have been widely used in creating special effects for television commercials, music videos, and movies. This thesis introduces new approaches to image morphing in three aspects: feature correspondence specification, warp generation, and transition behavior control. The approaches effectively solve the problems in each aspect and constitute an image morphing technique which overcomes the drawbacks of previous techniques. The features on two images are specified by points, line segments, and curves, and their correspondence is represented by a set of point pairs sampled on them. The tedium in specifying a feature is alleviated by adopting a computer vision technique called snakes. We consider a warp as the deformation of a plate in two dimensions and present two deformation techniques for deriving warps. The first is an energy minimization method and generates a $C^1$-continuous and one-to-one warp. The second method presents the multilevel free-form deformation as an extension of the free-form deformation and obtains a $^2$-continuous and one-to-one warp. Transition behavior control is separated from the feature interpolation in this thesis. The separation results in a method which is much easier to use than the previous approaches and enables the transition behavior to be specified by a function of time and positions. As extensions, we present a method for morphing between more than two images. And, the techniques for accelerating the generation of distorted images and for facilitating morphing between two image sequences are given.