Abstract

Vector graphics is broadly used in a variety of forms, such as illustrations, logos, posters, billboards, and printed ads. Despite its broad use, many artists still prefer to draw with pen and paper, which leads to a high demand of converting raster designs into the vector form. In particular, line drawing is a primary art and attracts many research efforts in automatically converting raster line drawings to vector form. However, the existing methods generally adopt a two-step approach, stroke segmentation and vectorization. Without vector guidance, the raster-based stroke segmentation frequently obtains unsatisfying segmentation results, such as over-grouped strokes and broken strokes. In this paper, we make an attempt in proposing an end-to-end vectorization method which directly generates vectorized stroke primitives from raster line drawing in one step. We propose a Transformer-based framework to perform stroke tracing like human does in an automatic stroke-by-stroke way with a novel stroke feature representation and multi-modal supervision to achieve vectorization with high quality and fidelity. Qualitative and quantitative evaluations show that our method achieves state of the art performance.

BibTex:

@inproceedings{liu2022-e2e-line-vec,
    author    = {Hanyuan Liu and Chengze Li and Xueting Liu and Tien-Tsin Wong},
    title     = {End-to-End Line Drawing Vectorization},
    booktitle = {Proceedings of Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI2022)},
    volume    = {36},
    number    = {4},
    year      = {2022},
    month     = {Feburary},
    pages     = {4559-4566}
}