Non-linear perceptual multi-scale network for single image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) show promise in single image super-resolution (SISR).
• Existing SISR networks often fail to fully utilize multi-scale features or use inefficient linear fusion methods.
• This leads to redundant feature extraction and suboptimal reconstruction performance.

Purpose of the Study:

• To propose a novel non-linear perceptual multi-scale network (NLPMSNet) for SISR.
• To address the limitations of single-stream and linear fusion approaches in existing SISR methods.
• To enhance the fusion of multi-scale image information for improved super-resolution.

Main Methods:

• Developed a novel non-linear perceptual multi-scale module (NLPMSM) utilizing a high-order channel attention mechanism.
• Introduced a multi-cascade residual nested group (MC-RNG) structure for capturing non-local hierarchical context.
• Employed a local residual nesting mechanism within LRNGs to stack NLPMSMs for effective residual learning.

Main Results:

• The proposed NLPMSNet demonstrates superior performance compared to state-of-the-art SISR methods.
• Achieved significant improvements in both quantitative metrics and visual quality.
• Maintained high performance with a notably small number of parameters.

Conclusions:

• NLPMSNet effectively fuses multi-scale image information non-linearly.
• The network architecture successfully captures discriminative multi-scale feature correlations and hierarchical context.
• NLPMSNet offers an efficient and high-performing solution for single image super-resolution tasks.