Transforming Image Super-Resolution: A ConvFormer-Based Efficient Approach.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Single-image super-resolution (SISR) methods face computational challenges for deployment on resource-limited devices.
• Transformer-based SISR models offer breakthroughs but incur significant computational overhead.
• Existing lightweight SISR methods struggle to balance performance and efficiency.

Purpose of the Study:

• To develop an effective and efficient lightweight SISR method.
• To address the high computational cost of transformer-based SISR models.
• To propose a novel architecture that combines convolutional and transformer advantages.

Main Methods:

• Introduced the Convolutional Transformer layer (ConvFormer) replacing self-attention with large kernel convolutions.
• Developed a ConvFormer-based Super-Resolution network (CFSR) for lightweight SISR.
• Proposed an edge-preserving feed-forward network (EFN) for local feature aggregation and high-frequency information preservation.

Main Results:

• CFSR demonstrates an optimal balance between computational cost and performance for lightweight SISR.
• CFSR achieved a 0.39 dB gain on the Urban100 dataset (x2 SR) compared to ShuffleMixer.
• CFSR requires 26% fewer parameters and 31% fewer FLOPs than comparable state-of-the-art methods.

Conclusions:

• CFSR offers an efficient and effective solution for lightweight single-image super-resolution.
• The proposed ConvFormer layer successfully models long-range dependencies with minimal overhead.
• CFSR provides a practical alternative for deploying high-performance SISR on resource-constrained devices.