Selective image abstraction

We present a novel and convenient method for producing selective stylized simplification of images. The user uses a brush to interactively mark certain areas of the input image which are to be left unaltered. Boundaries of these areas are then automatically optimized to underlying object boundaries in the image. Our method then performs stylized simplification of the unmarked areas, while preserving the marked areas. The method ensures a smooth transition between stylized and unaltered regions to leave a mixed reality image which combines the real and the abstract. Stylized simplification is performed using nonlinear diffusion, which can generate sophisticated results. We modify the classic model of nonlinear diffusion to incorporate bilateral filtering; we apply diffusion speed control of each pixel based on the user's input. The level of simplification can be controlled intuitively based on the diffusion time; another parameter controls the abstraction style, giving a simple and intuitive user interface. Our contributions include a simple-to-use method to produce a novel NPR style and a modified nonlinear diffusion model suited to this selective stylized simplification task. Experimental results show that the final mixed reality results are harmonious.