CSINet: A Cross-Scale Interaction Network for Lightweight Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep Convolutional Neural Networks (CNNs) have advanced image super-resolution (SR).
• Increasing CNN depth and breadth improves performance but raises computational and memory costs, limiting practical use.
• Efficient SR models are needed to balance performance and resource utilization.

Purpose of the Study:

• To develop a lightweight yet effective deep learning network for image super-resolution.
• To reduce the computational complexity and memory footprint of SR models.
• To improve the practical applicability of SR technology.

Main Methods:

• Incorporated factorized convolution and introduced the Cross-Scale Interaction Block (CSIB).
• CSIB utilizes a dual-branch structure (local and global features) with intermediate interaction for cross-scale information integration.
• Designed Efficient Large Kernel Attention (ELKA) with large kernels and gating for refining contextual information.

Main Results:

• Developed CSINet, a lightweight cross-scale interaction network for image super-resolution.
• CSINet significantly reduces computational costs while maintaining high performance.
• CSINet-S achieved state-of-the-art results on lightweight SR benchmarks with minimal parameters (e.g., 33.82 dB@Set14 × 2 with 248K parameters).

Conclusions:

• CSINet offers an efficient solution for practical image super-resolution applications.
• The proposed method outperforms existing lightweight SR techniques.
• CSINet demonstrates the effectiveness of cross-scale interaction and efficient attention mechanisms in SR.