Effective Clipart Image Vectorization through Direct Optimization of Bezigons.

Area of Science:

• Computer Vision
• Image Processing
• Computational Geometry

Background:

• Bezigons (Bezier curves) are crucial for shape representation in image vectorization.
• Existing methods often approximate intermediate vector formats, leading to errors and imperfections, especially in low-resolution images.
• Lack of curve priors and fitting ambiguities further challenge current vectorization techniques.

Purpose of the Study:

• To develop a novel method for vectorizing clipart images by directly optimizing Bezier curves (bezigons).
• To achieve higher fidelity and more reasonable bezigons compared to existing methods.
• To enhance the efficiency and quality of image vectorization.

Main Methods:

• Direct optimization of bezigons, bypassing intermediate representations.
• Development of a differentiable data energy term.
• Incorporation of curve-based prior terms for improved fitting.
• Utilization of an overlapped piecewise optimization strategy leveraging the local control property of bezigons.

Main Results:

• The proposed method generates bezigons with higher fidelity to the original raster image.
• Resultant bezigons are more reasonable and exhibit improved quality.
• Experimental results demonstrate superior performance compared to state-of-the-art methods and commercial software.

Conclusions:

• Direct bezigon optimization offers a more effective approach to image vectorization.
• The novel method addresses limitations of previous techniques, yielding superior results.
• This work advances the field of image vectorization, particularly for clipart and low-resolution imagery.