Saliency-preserving Image Warping

Developing an easy-to-use image warping technique and generating visual pleasing results have long been a hot research topic in the image processing field, which is valuable in both animation and image modeling field. In this paper we first analyze several popular image warping algorithm, then we propose a novel non-homogeneous saliency preserving image warping algorithm.The existing image warping algorithms can be categorized into pixel-based image warping based on scattered data-point interpolation methods and mesh-based texture mapping based on geometrical transformations. In the aspect of scattered data interpolation, several algorithms are reviewed in this paper such as the inverse distance weight (IDW), radial-based function (RBF) and moving least square (MLS). On the other hand, several algorithms based on triangular mesh and quadrilateral mesh are also reviewed. In the aspect of user interaction approach, those algorithms can be categorized into point controlling, curve controlling and target boundary specifying approaches. We have implemented these algorithms and make an intensive comparison among them.For mesh-based warping algorithms, first a mesh is embeded in the image domain, then it is rescaled or deformed based on the image and specified boundary constraints, finally the image pixels are mapped into the new image according to the deformed mesh. The existing warping methods tend to preserve the salient region of image during the warping by minimizing various mesh deformation energies.In this paper we propose a novel quadrilateral mesh-based image warping algorithm, which considers the image importance covered by each quad and distribute the distortion and rotation of the warping rationally onto every single quad in order to protect those portion with high image saliency. We properly reformulate the warping energy in a reduced way such that the result can be converged quickly during the iteration. Lots of experimental results have shown that our method is a novel and efficient image warping method which can well preserve the salient features of the input images.