A Real-World Animation Super-Resolution Benchmark With Color Degradation and Multi-Scale Multi-Frequency Alignment

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Animation super-resolution (SR) aims to enhance low-resolution (LR) animation frames to high-resolution (HR).
• Existing SR methods often fail with animation-specific degradations, differing from real-world image issues.
• Digital animations face unique compression and transmission artifacts like color banding and shifts.

Purpose of the Study:

• To introduce a novel benchmark, ADASR, for real-world animation super-resolution.
• To propose a new method, Color-Aware Animation Super-Resolution (CAASR), addressing animation-specific degradations.
• To improve the restoration of high-resolution frames for both 2D and 3D animations.

Main Methods:

• Developed the ADASR dataset featuring 2D and 3D animation content.
• Proposed CAASR, incorporating a color degradation simulation tailored for animations.
• Implemented a multi-scale, multi-frequency alignment mechanism for robust feature extraction.

Main Results:

• CAASR effectively addresses color banding, blocking, and color shifts in animation frames.
• The method achieves state-of-the-art performance on the AVC dataset and the new ADASR dataset.
• Demonstrated superior restoration of HR frames for both 2D and 3D animation content.

Conclusions:

• The proposed ADASR benchmark and CAASR method significantly advance animation super-resolution.
• CAASR's color-aware approach and robust feature extraction are key to its success.
• The findings facilitate industry applications requiring high-quality animation frame restoration.