Research On Multi-levels Object Based Image Vectorization

Image vectorization is the process of converting a raster image into a vector graphics. Raster images are understood by a computer only as an array of pixel values, limiting the operations on objects. Vector graphics, on the other hand, having a geometric data structure such as points, lines and faces associated with them, offer a more compact representation, resolution-independence, and geometric editability. Unfortunately, vector graphics are very time consuming and more difficult to produce than photographs, usually requiring a graphic artist to laboriously hand craft them. Through image vectorization, we can quickly convert raster images into vector graphics with minimal user intervention.However, most existing methods work on the whole image or single object, which can not represent different objects with their interactions between each other. Another difficulty is representing the subtile color change with a simple vector structure. In this dissertation, we resolve the problems above by discussing the issues of multi-level objects structure establishment, semantic relationships between different objects, multi-level objects abstraction in images and video sequences, topology-preserving objects surface mapping and texture-preserving vectorization.The main contributions of this thesis are as follow.1) Multi-level objects structure is established with four basic semantic relationships betweeen different objects to represent their interactions. This multi-level structure works on the whole process of image vectorization and affects every section as their foundation.2) A multi-level image objects abstraction method is proposed with their semantic relationships to correct the mis-labeled area. The using of semantic relationships between objects makes the rusults robuster and more accurate. 3) A novel method of video objects tracking and abstraction is proposed which could handle object partial occlusion and deformation, especially non-rigid object by considering an non-rigid object as the combination of rigid sub-objects.4) A novel approach of surface mapping is proposed to parameterize the multi-level objects by preserving their relationships and topology for the subsequent step of vectorization.5) We use a centroidal Voronoi tessellation method to generate a trigonal mesh steered by an image density map which keeps the color variety information. The proposed method represents the subtile color change with a simple vector structure.6) We subdivide the mesh grids with the information of vision attention to make the representation more accurate for important details.All those tehniques can be widely used in the domain of image understanding, image editing, key-frame animation, flash based digital game, digital entertainment, photorealistic/non-photorealistic rendering and industrial design. Widely applications make the research significative and valuable.