PNEN: Pyramid Non-Local Enhanced Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Traditional neural networks for image processing use limited kernel sizes, restricting contextual information capture.
• Exploiting long-range dependencies in images is computationally challenging with existing methods.

Purpose of the Study:

• To propose a novel non-local module, the Pyramid Non-local Block, for efficient exploitation of pairwise dependencies across different scales.
• To enhance pixel-level feature representation by learning correlations between multi-scale features.

Main Methods:

• Developed a Pyramid Non-local Block module that connects every pixel with all other pixels.
• Implemented a query feature map at full resolution and pyramid reference feature maps at downscaled resolutions.
• Exploited correlations between multi-scale reference features for enhanced representation.

Main Results:

• Achieved state-of-the-art performance in edge-preserving image smoothing by imitating classical algorithms.
• Demonstrated consistent performance improvements when integrating the Pyramid Non-local Block into existing image denoising and super-resolution methods.
• Showcased the module's efficiency in terms of memory and computational cost.

Conclusions:

• The Pyramid Non-local Block effectively captures long-range dependencies and enhances feature representation in low-level image processing.
• The module offers a computationally economical approach for improving various image restoration tasks.
• The Pyramid Non-local Block is versatile and can be integrated into existing convolutional neural networks for diverse applications.