Automatic Blending of Multiple Perspective Views for Aesthetic Composition

Hand-drawn pictures differ from ordinary perspective images in that the entire scene is composed of local feature regions each of which is projected individually as seen from its own vista point. This type of projection, called nonperspective projection, has served as one of the common media for our visual communication while its automatic generation process still needs more research. This paper presents an approach to automatically generating aesthetic nonperspective images by simulating the deformation principles seen in such hand-drawn pictures. The proposed approach first locates the optimal viewpoint for each feature region by maximizing the associated viewpoint entropy value. These optimal viewpoints are then incorporated into the 3D field of camera parameters, which is represented by regular grid samples in the 3D scene space. Finally, the camera parameters are smoothed out in order to eliminate any unexpected discontinuities between neighboring feature regions, by taking advantage of image restoration techniques. Several nonperspective images are generated to demonstrate the applicability of the proposed approach.