Reference-Based Deep Line Art Video Colorization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Art

Background:

• Manual line art coloring in animation is labor-intensive and time-consuming.
• Existing methods struggle with style consistency and temporal coherence in video coloring.

Purpose of the Study:

• To develop a deep architecture for automatic line art video coloring.
• To match the color style of reference images accurately.
• To ensure temporal color consistency in animated sequences.

Main Methods:

• A novel deep architecture combining a color transform network and a temporal refinement network (3U-net).
• A distance attention layer for region correspondence and local color transfer.
• Adaptive Instance Normalization (AdaIN) for global color style consistency.
• 3D convolutions for spatiotemporal feature learning.

Main Results:

• The proposed method achieves superior performance in line art video coloring.
• The model effectively transfers color styles from reference images.
• It ensures high temporal color consistency across video frames.
• The model demonstrates adaptability to new animation styles with fine-tuning.

Conclusions:

• The deep architecture offers an efficient and effective solution for automatic line art video coloring.
• The distance attention and AdaIN mechanisms successfully address style and region matching challenges.
• The approach significantly advances the state-of-the-art in automated animation coloring.