Global Learnable Attention for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Self-similarity is key in single image super-resolution (SISR).
• Existing methods often prioritize high-similarity non-local textures.
• This can limit detail recovery in severely damaged image regions.

Purpose of the Study:

• To challenge the assumption that high similarity always equates to importance in SISR.
• To develop a method that leverages low-similarity, yet informative, non-local textures.
• To improve the accuracy and richness of details in super-resolved images, especially under severe degradation.

Main Methods:

• Proposed Global Learnable Attention (GLA) to adaptively adjust similarity scores.
• Introduced Super-Bit Locality-Sensitive Hashing (SB-LSH) for efficient GLA computation.
• Integrated GLA as a general building block into deep SISR models, forming a Deep Learnable Similarity Network (DLSN).

Main Results:

• Demonstrated that low-similarity textures, often due to scale or orientation differences, can provide superior details.
• Achieved significant performance improvements in SISR across various degradation types (blur, noise).
• Reduced computational complexity of texture search from quadratic to near-linear using SB-LSH.

Conclusions:

• GLA effectively utilizes informative low-similarity textures for superior SISR.
• SB-LSH enables efficient application of GLA for large images.
• The proposed DLSN sets a new state-of-the-art in SISR, offering a versatile approach.