Towards object-based image editing.

With the increasing use of images in web design, document processing, entertainment, medical analysis, virtual environment creation, etc., the demand is dramatically growing for effective editing techniques that can fast and accurately create, compose, render and manipulate image contents. In recent years, plenty of research has been conducted for these tasks. However, the current techniques for these tasks are still far away from being satisfactory. It usually needs extensive user guidance, with painstaking time and effort, to produce a desired result. Our research thus investigates new techniques and tools for effective creation, extraction, composition, and other manipulations of image contents. We introduce an object oriented and vector based image editing framework. With the framework, we perceive an image as a set of objects represented by vector graphics so that image editing can be performed easily and semantically. The framework consists of four technical components: image segmentation, shape completion, image completion and image vectorization. The first three components are used to decompose an image into meaningful objects to support object-level editing. The last component is used to convert a raster image or an object of the image into a vector graphics representation to facilitate editing process. We have developed new algorithms for these components to implement the proposed framework. In particular, to overcome the lack of effectiveness and accuracy in extracting objects from complex backgrounds (especially with textures or low contrasts), we propose to incorporate image texture information into energy functions in graph-based image segmentation, use structure tensors to modify the weight between two nodes in the image graph, adopt a constrained active contour model to handle region and boundary simultaneously, and design a ``soft" brush interface for users to locally adjust segmentation results. The experiments show that our proposed method is robust to user inputs and is able to produce more accurate and smoother boundaries when evaluated on the MSRC image segmentation benchmark.DOCTOR OF PHILOSOPHY (SCE