| Although the acquisition and generation of digital images and videos have become more convenient than ever before, editing these materials remains very demanding for ordinary users. Despite the development of off-the-shelf image editing tools like Photoshop and Instagram, many high-level editing tasks, such as compositing and stylization, are still extremely challenging for non-professional users. These difficulties stem from two key factors. First, a (usually large) number of low-level operations are combined to achieve a high-level editing effect. Consequently, identifying the necessary operations may be an insurmountable obstacle for ordinary users. Second, the objectives of these editing tasks are often vaguely defined, usually with only subjective evaluations, such as "the composite looks realistic/fake," or "that film has a blue color tone, making me depressed." As a result, users with less domain knowledge have to resort to trial-and-error until they achieve the desired effect, which can be a time- and effort-intensive process.;In this thesis, we explore the automatic techniques to aid in high-level image editing, while minimizing users' required expertise. Our critical insights are 1) the use of machine learning to gain knowledge from massive datasets of images and videos; and 2) the establishment of perceptual metrics to quantitatively evaluate the effectiveness of editing. We conducted a series of studies to demonstrate that these insights help resolve the two aforementioned causes of difficulties and thus accomplish high-level image editing tasks.;First, we establish simple models of colors for familiar objects such as sky; grass, and skin, taking advantage of crowd-sourced perceptual studies. These identified memory colors are shown to be simple, yet effective, tools in enhancing image color reproduction. Second, we explore more complicated models of colors that are used to convey certain styles (such as genre, emotion, and directors) in feature films. By analyzing a collection of labeled film clips, we extract the low-level features associated with distinctive color styles and then apply them to stylize homemade images and videos. Third, we study the problem of color compatibility when combining parts of two different images, i.e. image compositing. Using both statistical and perceptual experiments, we identify critical features and their impacts on color compatibility. We propose a machine learning based technique to automatically adjust a composite for realistic colors. And, last but not least, we explore a very challenging image editing task that manipulates attributes beyond "color", such as texture and geometry. We demonstrate that by analyzing data collected from the target image itself and from external image resources, challenging weathering effects, such as stone erosion in a single photograph, can be compellingly simulated. |
