Towards Lightweight Super-Resolution With Dual Regression Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep neural networks excel at image super-resolution (SR) but face ill-posed problems.
• Existing SR methods struggle with large mapping spaces and computationally expensive large models.
• Model compression is challenging due to difficulties in identifying redundancy within vast SR mapping spaces.

Purpose of the Study:

• To reduce the ill-posed nature of SR by constraining the mapping space.
• To develop computationally efficient SR models without sacrificing performance.
• To propose a novel compression method for SR models.

Main Methods:

• A dual regression learning scheme is proposed, adding a secondary mapping to estimate downsampling kernels and reconstruct low-resolution (LR) images.
• This dual mapping constrains the SR mapping space, mitigating the ill-posedness.
• A dual regression compression (DRC) method is introduced for layer-level and channel-level compression using channel pruning.

Main Results:

• The dual regression approach effectively reduces the space of possible SR mappings.
• The DRC method successfully identifies and prunes redundant components in SR models.
• Experiments demonstrate the creation of accurate and efficient SR models.

Conclusions:

• The proposed dual regression learning scheme and DRC method effectively address key challenges in image super-resolution.
• This approach yields accurate and computationally efficient SR models.
• The findings contribute to advancing the field of deep learning for image restoration tasks.