WHANet: Weighted Hierarchical Attention for Lightweight SR

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Current Transformer models for image super-resolution (SR) often lose effective features due to sequential module arrangements.
• Local window-based feature extraction in Transformers can lead to the loss of crucial global information.

Purpose of the Study:

• To propose a novel model, Weighted Hierarchy Aggregation (WHA), to address feature loss in image SR.
• To enhance the capture of both local and global features without increasing model complexity.

Main Methods:

• Introduced the Hierarchy Aggregation Block (HAB) with learned weighting for adaptive feature emphasis.
• Developed the Global Residual Self-Attention Block (GRSAB) incorporating 2D channel attention to preserve global features.

Main Results:

• The WHA model accelerates network training and significantly improves SR performance.
• Achieved a 26.88dB PSNR score on the Urban100 dataset for ×4 SR tasks.
• Demonstrated high efficiency of HAB, particularly in lightweight SR applications.

Conclusions:

• WHA effectively mitigates feature loss in image SR by integrating HAB and GRSAB.
• The proposed model offers a superior approach to capturing multi-scale features for enhanced image reconstruction.
• WHA shows strong potential for practical applications in image super-resolution.