Deeply Recursive Low- and High-Frequency Fusing Networks for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) have significantly advanced single image super-resolution (SISR).
• Deeper and wider networks improve image reconstruction but increase computational and memory demands.
• Existing SISR models face challenges with training complexity and prediction time due to large network sizes.

Purpose of the Study:

• To propose a novel deeply-recursive low- and high-frequency fusing network (DRFFN) for efficient SISR.
• To address the trade-off between network size and performance in deep learning-based SISR.
• To enhance feature representation and image reconstruction quality with reduced resource consumption.

Main Methods:

• Developed a compact deeply-recursive network (DRFFN) utilizing parallel branches for low- and high-frequency information extraction.
• Incorporated a variance-based channel-wise attention (VCA) mechanism to optimize feature map information distribution.
• Employed cascading recursive learning with shared weights across convolutional recursions for model compactness.

Main Results:

• The DRFFN model demonstrated superior performance compared to existing SISR models on standard benchmark datasets.
• Achieved competitive quantitative and visual results in image super-resolution tasks.
• The compact network structure (DRFFN and DRFFN-L) offers significant advantages in computational and memory efficiency.

Conclusions:

• The proposed DRFFN effectively balances image reconstruction quality and computational efficiency for SISR.
• The parallel branch structure and VCA mechanism contribute to enhanced feature extraction and information fusion.
• DRFFN presents a promising solution for practical applications requiring high-quality image super-resolution with limited resources.