Image Retargeting Via Adaptive Scaling With Geometry Preservation

We present a novel grid-based image retargeting approach that is able to balance important/unimportant contents while minimizing visual distortions on image geometric structures. Our approach begins with a structure-consistent importance map, efficiently obtained by filtering image saliency under the guidance of image gradients. Then we maintain the important contents located by the importance map, and more importantly balance the spatial distribution adaptively between the important/unimportant contents by globally optimizing a shape parameter to constrain scaling factors. Moreover, to preserve some geometric structures (e.g., straight lines and circles) which occupy multiple quads, a Laplacian regularization term is exploited to smoothly propagate distortions. Finally, all these constraints are cast into a quadratic programming, the global optima of which can be found efficiently. Both subjective and objective evaluations are conducted to show the effectiveness of our approach.