Laplacian-based image retargeting with salient object magnified

Abstract

A compelling image retargeting approach aims at maintaining the important contents in the original image as much as possible, while minimizing the visual distortions on the image geometric structures. In this paper we present such an approach that succeeds in preserving the shape and size of the important contents and the geometric structures. We first encourage salient objects to be magnified by handling the scaling factor according to the corresponding saliency map. However, this may lead to over-magnification and distort the geometric structures (e.g. straight line and curve). We solve these problems by enforcing the deformed Laplacian coordinate of each grid vertex to be close to its original version, by which dramatic changes among neighboring quads are smoothed. All these constraints are integrated into a quadratic programming. The performance of the proposed approach is demonstrated by comparing with the state-of-the-art methods.