SAT: shift alignment transformer for video denoising without flow estimation.

Area of Science:

• Computer Vision
• Signal Processing
• Artificial Intelligence

Background:

• Video denoising seeks to restore clean video sequences from noisy frames.
• Existing methods struggle with long-range temporal information and complex motion.
• Transformer models face limitations with fixed windows for large-scale motion.

Purpose of the Study:

• To propose a novel and efficient Shift Alignment Transformer (SAT) for video denoising.
• To enable implicit feature alignment and aggregation over extended spatiotemporal regions.
• To overcome limitations of existing methods in handling complex motion and long-range dependencies.

Main Methods:

• Introduced Time Segment Shift for inter-frame correlation via dynamic temporal window shifting.
• Implemented Local Window Shift for enhanced intra-frame contextual modeling via spatial window displacement.
• Developed SAT to enable flexible receptive field expansion while maintaining computational efficiency.

Main Results:

• SAT consistently outperforms or matches state-of-the-art methods in denoising accuracy.
• Demonstrated effectiveness on both synthetic (RGB) and real-world (RAW) video denoising.
• Achieved a favorable balance between performance and computational cost.

Conclusions:

• SAT offers an effective and practical solution for long-range spatiotemporal modeling in video denoising.
• The proposed method provides a robust alternative to explicit motion estimation.
• SAT addresses limitations of fixed-window Transformer architectures for complex motion handling.