Fast and Robust Cascade Model for Multiple Degradation Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single Image Super-Resolution (SISR) is a challenging computer vision task.
• Existing methods often focus on simple blurs like Gaussian, limiting real-world applicability.
• Real-world camera movements introduce complex, non-Gaussian blurs.

Purpose of the Study:

• To develop a robust SISR model capable of handling large, non-Gaussian blurs.
• To propose a novel CNN cascade architecture for improved deblurring and upsampling.
• To integrate domain knowledge at the module level within the CNN for task-specific constraints.

Main Methods:

• A new Convolutional Neural Network (CNN) cascade model is formulated.
• Each sub-module is specialized for either deblurring or upsampling.
• A densely connected CNN architecture incorporates external knowledge to constrain sub-module outputs.
• A final sub-module addresses residual errors from preceding modules.

Main Results:

• The proposed model successfully manages a wider range of deformations, including large non-Gaussian blurs.
• It outperforms current state-of-the-art (SOTA) methods on standard SISR datasets.
• The model demonstrates improved computational efficiency compared to close competitors.
• The approach shows robustness to errors in blur kernel estimation.

Conclusions:

• The novel CNN cascade effectively addresses limitations of existing SISR methods for real-world blurs.
• Domain knowledge integration at the module level is a key innovation for SISR.
• The model offers a promising alternative to blind SISR models due to its robustness and efficiency.