An adaptive cartoon-like stylization for color video in real time

The cartoon-like stylization of color video images is one of the major types of non-photorealistic rendering. It has broad applications in the entertainment industry due to its lively style, flexible form, and autonomous creation. Our video cartoonization framework consists of color space conversion, bilateral filtering, color quantization and edge detection. We use an adaptive difference of Gaussians filter to detect the edges– variance value of Gaussian filter is determined based on the gradient of each pixels, so that we can not only detect elaborate edges but also decrease the influence of noise. We also alter the target sharpness range in the color quantization to achieve a softer look. We implement two versions of our algorithm: serial on a CPU, and parallel on a GPU having the Computer Unified Device Architecture (CUDA). Only the CUDA GPU code is able to process video in real time, running 45 to 180 times faster than the CPU code.