Lightweight local and global granularity selection optimization network for single image super-resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Single Image Super-Resolution (SISR) models benefit from combining local and global features.
• Existing methods often use linear fusion for local features, causing redundancy and inefficient extraction.
• Global feature extraction is hindered by irrelevant features, impacting dependency capture and reconstruction quality.

Purpose of the Study:

• To propose a lightweight network, LGGSONet, for enhanced feature extraction in SISR.
• To improve the efficiency and effectiveness of both local and global granularity feature processing.
• To enhance the overall reconstruction quality of SISR models.

Main Methods:

• Introduced a Local Granularity Selection Module (LGSM) using nonlinear convolution for dynamic multi-scale feature fusion.
• Developed a Global Granularity Optimization Module (GGOM) employing global transposed attention to filter irrelevant features.
• Integrated LGSM and GGOM into a Mixed Granularity Transformer Block (MGTB) and a Mixed Granularity Residual Transformer Group (MGRTG) for simplified network training.

Main Results:

• LGGSONet achieved a 0.30 dB PSNR improvement over advanced lightweight methods.
• The proposed network maintains fewer parameters and lower computational costs.
• Experimental results validate the effectiveness of LGGSONet in SISR tasks.

Conclusions:

• LGGSONet effectively addresses the limitations of linear feature fusion and irrelevant feature interference in SISR.
• The proposed network offers a lightweight and efficient solution for high-quality image super-resolution.
• The novel modules (LGSM, GGOM) and architecture (MGTB, MGRTG) significantly advance SISR capabilities.