Image Abstraction Based On Visual Attention Model And Shape-Simplifying

With the development of computer graphics, more and more people are used the different ways to simulate natural scenes. In some applications, for example, network video chat. We urgently need this technique to express and transfer information besides photorealistic rendering, such as, shape-simplifying, cartoon, and lines drawing the character and background of image, which depend on the scene to express the main information of the image, The abstraction rendering was made up for people needing, the abstracted visual information is better to attract the attention of people. So, visual information became a bridge for the exchange of information. Abstraction image is import technology which transfer photos into non-photorealistic effects images.This thesis introduces the research background, applications, technology status of abstraction. Then, we introduce our two parts of work, image abstraction based on visual attention model and shape-simplifying.This thesis first work is image abstraction based on visual attention model. First, in order to reduce the R, G, B three-channel correlation, we transform the input image from RGB color space to CIE-Lab color space. Second, we iteratively used bilateral filter to construct a smooth feature preserved edge tangent flow filed, while high contrast regions are strengthened with the flow-based difference of Gaussian filter, we get a map of lines drawing. Other, we adopt a more elaborated visual attention (saliency) model to get a region of interest function maps, the visual attention model is built based on human attention and reflect most people's attention. We based on this interest function maps to get a preliminary abstraction by bilateral filter, then tutors map of lines drawing add to preliminary abstraction image. Finally the soft luminance quantization is adopted to further enhance the cartoon-like effect. The result image transfer from the CIE-Lab color space to RGB color space.This thesis second work is tried to simplify the shape of the images, we use the feature flow field constraints on the mean curvature flow filter in the R, G, B three channels, respectively, iteration simplify and shrink the image shape, then use shock filter to protect important shape edges. The method has three parts, first, we iteratively used a bilateral filter to construct a smooth, and coherence feature preserved edge tangent flow filed, which indicates the salient feature directions of the image, smoothing no-salient feature directions of the image. Second, we iteratively simplify and shrink the shape of image with mean curvature flow. Finally, we iteratively using shock filter to protect important shape edges. Because the mean curvature flow not only simplify and shrink the image shape, but also can simplify the color. So, we don't need other processes to do.