Real-World Light Field Image Super-Resolution Via Degradation Modulation.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Deep neural networks (DNNs) have advanced light field (LF) image super-resolution (SR).
• Existing DNN-based LF-SR methods are limited to single, fixed degradations, failing on real-world images with varied quality.
• There is a need for LF-SR methods robust to diverse, real-world image degradations.

Purpose of the Study:

• To develop a novel method for real-world LF image super-resolution (SR) that addresses limitations of existing approaches.
• To create a practical LF degradation model for formulating real-world LF image degradation processes.
• To design a convolutional neural network (CNN) capable of handling diverse degradations while preserving spatial and angular information.

Main Methods:

• A practical LF degradation model was developed to simulate real-world LF image degradation.
• A CNN was designed to integrate degradation priors into the SR process.
• The network was trained on LF images using the formulated degradation model, enabling it to adapt to various degradation types.

Main Results:

• The proposed method demonstrated superior SR performance on both synthetically degraded and real-world LF images.
• The method achieved better generalization capabilities on real-world LF images compared to state-of-the-art techniques.
• Experimental results confirmed the effectiveness of the network in modulating different degradations while utilizing spatial and angular information.

Conclusions:

• The developed method offers an effective solution for real-world LF image super-resolution across a wide range of degradations.
• The approach generalizes well to real-world LF images, outperforming existing single and LF image SR methods.
• The proposed LF degradation model and CNN architecture provide a robust framework for tackling diverse LF image quality issues.