RepECN: Making ConvNets Better Again for Efficient Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Traditional Convolutional Neural Network (CNN)-based image super-resolution (SR) methods offer low computational costs but limited performance.
• Vision Transformer (ViT)-based SR methods achieve high performance but suffer from significant computational and storage overhead.
• Practical SR applications demand both high-quality reconstruction and fast inference.

Purpose of the Study:

• To develop a novel CNN-based model for image super-resolution that balances high performance with computational efficiency.
• To address the limitations of existing SR methods in real-world scenarios requiring fast inference and minimal overhead.

Main Methods:

• Propose RepECN (Efficient Residual ConvNet enhanced with structural Re-parameterization), a CNN-based SR model.
• Employ a stage-to-block hierarchical architecture inspired by ViT, replacing Multi-Head Self-Attention (MHSA) with larger kernel convolutions in Re-Parameterization ConvNet Blocks (RepCNB).
• Utilize a novel image reconstruction module with nearest-neighbor interpolation and pixel attention, alongside bicubic interpolation for high-frequency information learning.

Main Results:

• RepECN achieves 2.5x to 5x faster inference speeds compared to state-of-the-art ViT-based SR models.
• The proposed model demonstrates competitive or superior super-resolving performance across multiple public benchmarks.
• RepECN effectively balances reconstruction quality and inference efficiency.

Conclusions:

• RepECN offers a superior trade-off between performance and efficiency for image super-resolution tasks.
• The model's design overcomes the computational burdens of ViT-based methods, making it suitable for practical applications.
• RepECN provides a viable solution for high-quality, fast image super-resolution.