ClipGen: A Deep Generative Model for Clipart Vectorization and Synthesis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computer Graphics

Background:

• Manual vectorization of raster clipart is time-consuming and requires specialized skills.
• Existing automated methods often struggle with complex shapes and maintaining visual fidelity.
• The demand for easily editable and scalable vector graphics in design is high.

Purpose of the Study:

• To develop a novel deep learning-based approach for automatic vectorization and synthesis of man-made object clipart.
• To generate layered vector clipart from raster images, preserving object category and visual characteristics.
• To create a generative model capable of producing human-recognizable vector graphics.

Main Methods:

• An iterative generative model was employed to sequentially generate new layers, each with a single color and a closed path.
• A joint loss function was formulated, incorporating shape similarity, symmetry, local curve smoothness, and rendering accuracy.
• A new dataset, ClipNet, comprising man-made object clipart with closed-path layers, was introduced alongside preprocessing tasks.

Main Results:

• The proposed method successfully vectorized and synthesized clipart across various man-made object categories.
• Experimental validation demonstrated the model's ability to generate recognizable and accurate vector graphics.
• The generated vector clipart maintained key characteristics such as shape similarity and symmetry.

Conclusions:

• The deep learning approach offers an effective solution for automated clipart vectorization and synthesis.
• The generative model facilitates efficient and intuitive clipart design, benefiting both novice users and professionals.
• This work contributes to advancing automated graphic design tools through AI-powered vectorization.