A Multi-Scale Recursive Attention Feature Fusion Network for Image Super-Resolution Reconstruction Algorithm.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) show promise in single-image super-resolution (SISR).
• Existing CNNs struggle with underutilizing feature information and losing critical details.
• SISR remains a challenging task requiring enhanced feature extraction and fusion.

Purpose of the Study:

• To propose a novel network, the multi-scale recursive attention feature fusion network (MSRAFFN), for improved SISR.
• To address feature underutilization and detail loss in current SISR methods.
• To enhance the quality and visual effects of super-resolved images.

Main Methods:

• A three-part network architecture: shallow feature extraction, multi-scale recursive attention feature fusion, and reconstruction.
• Utilizing a multi-scale recursive attention feature fusion network block (MSRAFFB) with an attention mechanism for feature enhancement and fusion.
• Employing residual connections for cross-layer integration of image features and deconvolution for upsampling and high-frequency information extraction.

Main Results:

• The MSRAFFN effectively fuses multi-scale features and enhances channel features via attention.
• Cross-layer connections with residual learning integrate features across different levels.
• The reconstruction module sharpens high-resolution images, improving visual quality.

Conclusions:

• The proposed MSRAFFN demonstrates superior performance in SISR tasks.
• The network achieves better subjective visual effects and objective evaluation metrics than existing models.
• MSRAFFN offers a promising approach for enhancing image resolution with deep learning.