Hybrid attention optimized hierarchical multiscale transformer architecture for image super-resolution

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer architectures show promise in image super-resolution (SR).
• Existing models struggle with local feature modeling and feature representation, hindering fine detail restoration in high-resolution (HR) images.

Purpose of the Study:

• To address limitations in current Transformer-based SR models.
• To propose a novel architecture for improved local feature interaction and detail reconstruction.

Main Methods:

• Introduced the Hierarchical Multiscale Transformer Architecture (HMT) with Hybrid Attention.
• Developed the Fourier Mamba Synergy (FMSA) module for cross-domain feature collaboration and long-range dependency modeling.
• Incorporated Dynamic Convolutional Attention (DCA) for multi-scale feature enhancement and instance weight adaptation.

Main Results:

• HMT demonstrated superior performance, exceeding HGFormer by 0.07dB and 0.08dB in PSNR at a 4x scaling factor on BSD100 and Manga109 datasets.
• Achieved significant improvements in both objective metrics and visual quality on public datasets.
• The proposed architecture effectively captures fine-grained local feature interactions for clearer detail generation.

Conclusions:

• The HMT architecture offers enhanced capabilities for image super-resolution.
• It effectively resolves the issues of poor local feature modeling and limited representation in existing Transformer models.
• Results indicate strong potential for generating high-quality, detailed high-resolution images.