Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Segmentation-guided Layer-wise Image Vectorization with Gradient Fills

Hengyu Zhou, Hui Zhang, Bin Wang

TL;DR

Segmentation-guided layer-wise vectorization addresses the challenge of converting raster images into concise vector graphics with radial gradient fills. It extends gradient support in a model-free framework by progressively adding gradient-filled Bézier paths guided by a gradient-aware segmentation, initialized through segmentation-guided priors and optimized with a segmentation-guided loss and Xing regularization. The key contributions are the gradient-aware segmentation, segmentation-guided initialization, and the SG loss that improves color propagation inside paths, validated against three emoji/icon datasets and a user study, showing faster convergence and higher perceived quality than prior methods. The approach enables topology-preserving, editable vector graphics with gradients without dataset-specific training, increasing practicality for diverse raster inputs.

Abstract

The widespread use of vector graphics creates a significant demand for vectorization methods. While recent learning-based techniques have shown their capability to create vector images of clear topology, filling these primitives with gradients remains a challenge. In this paper, we propose a segmentation-guided vectorization framework to convert raster images into concise vector graphics with radial gradient fills. With the guidance of an embedded gradient-aware segmentation subroutine, our approach progressively appends gradient-filled Bézier paths to the output, where primitive parameters are initiated with our newly designed initialization technique and are optimized to minimize our novel loss function. We build our method on a differentiable renderer with traditional segmentation algorithms to develop it as a model-free tool for raster-to-vector conversion. It is tested on various inputs to demonstrate its feasibility, independent of datasets, to synthesize vector graphics with improved visual quality and layer-wise topology compared to prior work.

Segmentation-guided Layer-wise Image Vectorization with Gradient Fills

TL;DR

Segmentation-guided layer-wise vectorization addresses the challenge of converting raster images into concise vector graphics with radial gradient fills. It extends gradient support in a model-free framework by progressively adding gradient-filled Bézier paths guided by a gradient-aware segmentation, initialized through segmentation-guided priors and optimized with a segmentation-guided loss and Xing regularization. The key contributions are the gradient-aware segmentation, segmentation-guided initialization, and the SG loss that improves color propagation inside paths, validated against three emoji/icon datasets and a user study, showing faster convergence and higher perceived quality than prior methods. The approach enables topology-preserving, editable vector graphics with gradients without dataset-specific training, increasing practicality for diverse raster inputs.

Abstract

The widespread use of vector graphics creates a significant demand for vectorization methods. While recent learning-based techniques have shown their capability to create vector images of clear topology, filling these primitives with gradients remains a challenge. In this paper, we propose a segmentation-guided vectorization framework to convert raster images into concise vector graphics with radial gradient fills. With the guidance of an embedded gradient-aware segmentation subroutine, our approach progressively appends gradient-filled Bézier paths to the output, where primitive parameters are initiated with our newly designed initialization technique and are optimized to minimize our novel loss function. We build our method on a differentiable renderer with traditional segmentation algorithms to develop it as a model-free tool for raster-to-vector conversion. It is tested on various inputs to demonstrate its feasibility, independent of datasets, to synthesize vector graphics with improved visual quality and layer-wise topology compared to prior work.
Paper Structure (24 sections, 7 equations, 14 figures, 2 tables)

This paper contains 24 sections, 7 equations, 14 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Image Vectorization
  4. Image Topology
  5. Method
  6. Method Overview
  7. Gradient-aware Segmentation
  8. Segmentation-guided Initialization
  9. Loss Function with Segmentation-guided weight
  10. Experiments
  11. Implementation Details
  12. Datasets
  13. Noto Emoji
  14. Fluent Emoji
  15. Iconfont
  16. ...and 9 more sections

Figures (14)

  • Figure 1: Overview of our framework. U+1F61A from Noto Emoji noto is used for demonstration.
  • Figure 2: Comparisons on segmentation methods. (a) is the raster input. (b) is by LIVE under default settings clustering colors into 200 bins. (c) is also by LIVE but with the number of bins set to 30. (d) is segmented using the Mean-shift comaniciu2002mean algorithm. (e) is by our method.
  • Figure 3: Step-by-step outputs of our gradient-aware segmentation
  • Figure 4: With paths being progressively added, the overall difference between the current output and the raster input decreases, thus we apply Otsu's method otsu1979threshold to automatically determine the threshold for binarization. In this example, the red cheeks get segmented after more significant differences including eyes and brows are fitted, thanks to the dynamic threshold. The differences are normalized and mapped to grayscale.
  • Figure 5: Optimizing course for our method in comparison with UDF loss from LIVE. With our proposed segmentation-guided weight (SG Weight), the gradient fill is optimized to minimize the color error not only on the contour as UDF weight focuses on but also on colors inside the path, while excluding pixels occluded by eyes, brows, and the mouth.
  • ...and 9 more figures