Treffer: Content-aware image retargeting and its quality assessment

Ph.D. ; In the last decade, we have witnessed the fast development of mobile devices, which has imposed new demands on convenient image display. Compared with the various sizes and aspect ratios of display devices, digital images only have a limited number of alternatives of sizes. Consequently, the research on image retargeting, which adapts the source image to arbitrary sizes, is becoming ever more important. Unlike traditional methods such as linear scaling and cropping, content-aware image retargeting resizes an image under two objectives: preserve salient content and avoid annoying visual artifacts. This technique creates a series of images that fit different display devices. These algorithms either remove less important pixels or warp the images to emphasize important objects. To assess the performance of different retargeting algorithms, many recent works focus on the quality assessment of image retargeting. In contrast with traditional image quality assessment, the quality assessment of image retargeting should also estimate the visual artifacts introduced to retargeted images, such as information loss, discontinuities and shape distortion. Therefore, many previous methods first align the source and the retargeted images and define the overall similarity between aligned pixels or patches as the retargeting quality. The research on retargeting performance helps to choose the retargeted image of the highest quality and recommend it to users. ; In the first part of the thesis, our proposed image retargeting algorithms are introduced. Content-aware image retargeting can be roughly categorized into two classes: discrete and continuous. Discrete algorithms remove less important pixels or patches to generate retargeted images. In contrast, continuous methods determine an optimal mapping from source pixels to target pixels. We mainly focus on improving the performance of continuous retargeting algorithms. Previous continuous methods overlay uniform mesh grids on the source image and optimize the deformation of ...