CartoonLossGAN: Learning Surface and Coloring of Images for Cartoonization

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Artistic style transfer methods struggle with cartoonization due to differences in image characteristics.
• Existing methods fail to capture the smooth surfaces and distinct color palettes of cartoon images.

Purpose of the Study:

• To develop an effective generative adversarial network (GAN) for high-quality image cartoonization.
• To introduce a novel cartoon loss function that mimics sketching and coloring processes.

Main Methods:

• Proposed a compact GAN architecture by reusing the discriminator's encoder.
• Introduced a novel cartoon loss function to learn smooth surfaces and coloring characteristics.
• Developed an initialization strategy for stable and easier training with discriminator reuse.

Main Results:

• CartoonLossGAN successfully generates high-quality, fantastic cartoon-style images.
• The proposed method demonstrated superior performance compared to four representative artistic style transfer techniques.
• Experimental results validated the effectiveness of the cartoon loss and initialization strategy.

Conclusions:

• CartoonLossGAN offers a significant advancement in image cartoonization.
• The novel loss function and architecture effectively address the challenges of cartoon style transfer.
• The method provides a robust and stable solution for generating cartoon images.