Lightweight image super-resolution based multi-order gated aggregation network

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Transformer-based models show promise for image super-resolution (SR) but suffer from high computational costs due to self-attention mechanisms.
• Existing methods often struggle to balance performance, computational complexity, and visual quality in SR tasks.

Purpose of the Study:

• To propose an efficient and effective image super-resolution network, the Multi-Order Gated Aggregation Super-Resolution Network (MogaSRN).
• To reduce the computational burden of Transformer-based SR models while maintaining or improving performance.

Main Methods:

• Introduced MogaSRN, inspired by MogaNet, for low-level vision tasks.
• Employs spatial multi-order context aggregation and adaptive channel-wise reallocation using multi-layer perceptrons (MLPs).
• Maintains fixed resolution during deep feature extraction, unlike MogaNet, to optimize for SR.

Main Results:

• MogaSRN demonstrates significant improvements over state-of-the-art methods on five benchmark datasets.
• The model achieves superior visual quality and reconstruction accuracy.
• Achieved 3.7x faster runtime for 4x scaling compared to LWSwinIR with better performance.

Conclusions:

• MogaSRN offers a compelling balance between performance and model complexity for image super-resolution.
• The proposed architecture provides an efficient alternative to computationally intensive Transformer-based SR models.
• MogaSRN delivers high-quality image reconstruction with reduced computational overhead.