Feature-Aware Mesh for Image Retargeting

Abstract

Image retargeting adapts an input image to displays other than originally intended while minimizing visual distortion. Most warping-based methods formulate retargeting as a spatially-varying warping problem over a uniform mesh and encourage each mesh grid to undergo a conformal transformation to minimize visual distortion. However, the conformal transformation constraint on each individual mesh grid is often insufficient to preserve global image structures. In this paper, we present a feature-aware mesh based retargeting method. Our idea is to warp each visually salient image feature as a whole, thus preserving global image structures. Our method divides an image into a non-uniform mesh such that each salient image feature or object is enclosed only in one mesh element. Our method warps each mesh element with an affine transformation. We then formulate image retargeting as a spatially-varying affine warping problem. To further minimize the distortion, we encourage each affine transformation to be close to more distortion-free ones, such as conformal transformation. The spatially-varying affine warping problem is formulated as a quadratic energy minimization problem, which can be solved efficiently. Our experiments demonstrate that our method can effectively preserve global image structures.