Taylor Neural Network for Real-World Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Single image super-resolution (SR) research faces challenges due to synthetic data degradation.
• Real-world SR (RWSR) datasets reveal practical degradation, stressing deep neural network capabilities.
• Existing methods struggle with realistic image quality reconstruction.

Purpose of the Study:

• To address the limitations of synthetic degradation in SR research.
• To propose a novel deep neural network architecture for Real-World Super-Resolution (RWSR).
• To leverage Taylor series approximation for enhanced image reconstruction.

Main Methods:

• Development of Taylor Neural Networks (TNNs) based on Taylor series approximation.
• Introduction of Taylor Modules with Taylor Skip Connections (TSCs) to approximate feature projection.
• Integration of TSCs to aggregate high-order image details from different network layers.

Main Results:

• TNNs effectively approximate feature projection functions using Taylor series.
• TSCs enable attending to and aggregating diverse high-order image details.
• TNNs demonstrate compatibility with various neural network backbones with minimal parameter overhead.

Conclusions:

• TNNs offer a principled approach to RWSR by approximating complex image degradations.
• The proposed architecture enhances the learning of high-order image components for superior reconstruction.
• Extensive experiments confirm TNNs outperform existing methods on RWSR benchmarks.