Revitalizing Convolutional Network for Image Restoration.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Image restoration is crucial for reconstructing high-quality images from degraded versions.
• Transformer models have recently dominated image restoration due to their long-range pixel interaction capabilities.
• Convolutional Neural Networks (CNNs) were previously the standard but have been largely superseded by Transformers.

Purpose of the Study:

• To investigate the potential of CNNs for image restoration.
• To demonstrate that a simple CNN architecture can achieve competitive or superior performance compared to Transformer models.
• To identify key factors for improving image restoration model performance.

Main Methods:

• Development of a novel CNN architecture named ConvIR, utilizing efficient convolution operators.
• Re-examination of characteristics of advanced image restoration algorithms to inform network design.
• Extensive experimentation on 20 benchmark datasets across five image restoration tasks.

Main Results:

• The proposed ConvIR network performs comparably to or better than Transformer-based models.
• ConvIR achieves state-of-the-art performance in image restoration.
• The model demonstrates low computational complexity.

Conclusions:

• CNNs remain highly effective for image restoration, challenging the dominance of Transformers.
• ConvIR offers a computationally efficient and high-performing solution for various image restoration tasks.
• The findings provide insights into designing effective deep learning models for image restoration.