Abstract
This thesis presents a novel concept of interactive image editing: to edit an image by its attributes. We aim to alleviate the difficulty of conventional image editing methods, which often require the user to be experienced and skillful in manipulating the regions and contents of an image. We introduce an attribute-based image editing framework, and demonstrate two plausible editing tasks that can be effectively done using our framework. The first one is boundary adjustment with respect to the ‘semantic region size’ attribute and the second one is color transfer with respect to the ‘natural color’ attribute. We use the Fully Convolutional Network (FCN) model to segment a given image into several semantic regions and then characterize each region by the aforementioned attributes. For the boundary adjustment task, we adopt the seam carving method to adjust the semantic region size attribute of the selected region, and therefore the user is allowed to change the composition of the image interactively. For the color transfer task, we model the natural color attribute of the selected region by referring to the distribution of color attributes in the database. The experimental results show that our method is efficient and easy to generate visually pleasing editing results.