Topology- and perception-aware image vectorization (Q6186787)

In this paper, the authors introduce a novel vectorization method for converting color raster images into vector images. While previous approaches primarily focus on individually vectorizing polygons of different colors in the image, the authors' method emphasizes the structural properties of the polygons discernible in the image, enabling a systematic examination of both color and structural attributes. The method employs hierarchical analysis across three levels. At the lowest level, termed 'curve-elements,' the algorithm detects primitive elements. This is followed by the 'curvebases' level, where these primitives are assembled into higher-level structures. Finally, at the highest level, the image undergoes polygonal partitioning, forming the foundation for rendering within the application. A key advantage of the authors' method is its ability to operate stably regardless of the nature of the image, from photorealistic images to pixel art. Additionally, the method utilizes fewer Bézier curves during vectorization, even when producing high-quality vectors. The developed method (referred to as HCV) is benchmarked against algorithms from the literature and commercial software. The results demonstrate that high-quality vectorization can be achieved with a significantly reduced number of Bézier curves. Moreover, the proposed solution addresses an error present in some existing algorithms, which often results in gaps between individual regions of the images.