Attention Redundancy Reduction for Image Super-Resolution

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Transformer models show promise for Single Image Super-Resolution (SISR).
• Existing state-of-the-art (SOTA) models suffer from redundancy in attention maps, impacting quality and efficiency.
• High mutual information across attention maps is a key issue in current SISR transformer models.

Purpose of the Study:

• To propose a novel Low Redundancy Attention Network (LRAN) for efficient and high-quality SISR.
• To address redundancy issues within attention heads and across network blocks in transformer-based SISR.
• To improve the trade-off between image reconstruction quality and computational speed in SISR.

Main Methods:

• Introduced a multi-element mechanism in self-attention to increase inter-head diversity and mitigate head redundancy.
• Proposed an encapsulated architecture with enhanced local perception units and gated multi-layer perceptrons (MLPs) for improved local information capture.
• Integrated a single self-attention layer between multiple MLP layers within the encapsulated architecture.

Main Results:

• LRAN demonstrates superior performance compared to SOTA models in lightweight SISR tasks.
• LRAN achieves a better balance between image quality (e.g., PSNR) and processing speed.
• LRAN-light achieved 0.32dB PSNR higher than SwinIR-light on Urban100 for ×4 SR while being 4x faster.

Conclusions:

• The proposed LRAN effectively reduces redundancy in attention mechanisms for SISR.
• LRAN offers significant improvements in both quality and efficiency for lightweight SISR.
• LRAN presents a promising direction for developing more effective and efficient super-resolution models.