High-Similarity-Pass Attention for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Non-local attention (NLA) is crucial for self-similarity-based single image super-resolution (SISR).
• Standard NLA methods explore non-local self-similarity (NSS) but exhibit inefficiencies.
• Randomly selected regions in NLA yield similar performance to standard NLA, suggesting potential improvements.

Purpose of the Study:

• To analyze the attention map of standard NLA and identify its limitations in SISR.
• To develop a more efficient attention mechanism for SISR by focusing on high-similarity features.
• To introduce a novel attention block that can be integrated into existing deep SISR models.

Main Methods:

• Analysis of standard NLA attention maps to reveal statistical waste in feature assignment.
• Introduction of a soft thresholding operation to create high-similarity-pass attention (HSPA).
• Derivation of properties for end-to-end training of HSPA and integration into a deep high-similarity-pass attention network (HSPAN).

Main Results:

• HSPA generates a more compact and interpretable attention distribution by focusing on high-similarity features.
• The proposed HSPA can be seamlessly integrated as a general building block in deep SISR architectures.
• The HSPAN model, incorporating HSPA, achieved superior quantitative and qualitative results compared to state-of-the-art SISR methods.

Conclusions:

• The developed HSPA mechanism effectively addresses the inefficiencies of standard NLA in SISR.
• HSPA offers a more focused and efficient approach to modeling long-range dependencies in image super-resolution.
• The HSPAN model demonstrates significant advancements in SISR performance, validated by extensive experiments.