Real-time Image Vectorization on GPU

Abstract

In this paper, we present a novel algorithm to convert a raster image into its vector form. Different from the state-of-art methods, we explore the potential parallelism that exists in the problem and propose an algorithm suitable to be accelerated by the graphics hardware. In our algorithm, the vectorization task is decomposed into four steps: detecting the boundary pixels, pre-computing the connectivity relationship of detected pixels, organizing detected pixels into boundary loops and vectorizing each loop into line segments. The boundary detection and connectivity pre-computing are parallelized owing to the independence between scanlines. After a sequential boundary pixels organizing, all loops are vectorized concurrently. With a GPU implementation, the vectorization can be accomplished in real-time. Then, the image can be represented by the vectorized contour. This real-time vectorization algorithm can be used on images with multiple silhouettes and multiview videos. We demonstrate the efficiency of our algorithm with several applications including cartoon and document vectorization.

Vectorizing binary image boundaries with symmetric shape detection, bisection and optimal parameterization

“…Fitting quality quantified by fitting errors plays an important role in representing original image boundaries accurately. Due to its importance, existing vectorization methods mainly focus on accuracy during the conversion from raster images to vector images 33 . The proposed method has smaller fitting errors than existing methods, demonstrating its advantage in accurately representing original image boundaries.…”