Robust image retargeting via structure aware deformation

Abstract

Content-based image retargeting is a technique that resizes an input image to a given target resolution while minimizing distortions of important objects caused by aspect ratio variations. Conventional approaches share the similar methodology which aims to preserve salient regions as much as possible while allowing distortions of trivial regions. Those methods show satisfactory results for input images whose objects are distinct and backgrounds are monotonous. However, their performance is not always guaranteed for images having structural components such as straight lines, which are prone to be distorted after resizing and sensitive to human perceptions. In this thesis, we propose an image retargeting algorithm that is robust for structure preservation. Based on axis-aligned grid, it finds the optimal grid for target image by a quadratic optimization represented by the two objective functions. The first one is As-similar-as-possible (ASAP) energy function, which aims to preserve important regions while allowing distortions of trivial regions. The second one is Adaptive Laplacian regularization (ALR) energy function, which aims to relieve structural distortions. Those two energy functions are combined into single quadratic optimization model ensuring the global convexity and it is solved by quadratic programming solver for the optimal grid. Experimental results show that our method is robust for preserving structural components while achieving the basic purpose of content-based image retargeting. For more reliable comparisons with other methods, we demonstrate objective evaluation scores obtained by the image retargeting quality assessment scheme.