OMLK-Net: An Online Multi-scale Large Separable Kernel Distillation Network for efficient image super-resolution

Single-image super-resolution (SISR) has seen remarkable progress in recent years, driven by the powerful learning capabilities of large-scale neural networks, such as deep CNNs and Transformers. However, these advances come at the expense of substantial computational costs. Striking a delicate balance between effectiveness and efficiency remains a key challenge in neural network design. This paper proposes OMLK-Net, a novel lightweight architecture for SISR, offering the dual advantages of computational efficiency and high effectiveness. OMLK-Net adopts a divide-and-conquer strategy to separately optimize local and nonlocal feature learning, enabling a lightweight architecture without compromising feature representation effectiveness. Specifically, our OMLK-Net comprises two key modules: an Online Multiscale Distillation Block (OMDB) and Large Separable Shifting Kernel Attention (L2SKA) blocks. The OMDB module aims to explore multiscale local contextual information with a customized lightweight network block; while the L2SKA aims to harness nonlocal features by using computationally efficient large separable shifting kernels. By virtue of its carefully designed local and nonlocal feature extraction operators, OMLK-Net effectively addresses SISR challenges while maintaining low computational complexity. Extensive experimental results on benchmark datasets demonstrate that OMLK-Net achieves a better trade-off against state-of-the-art methods in terms of performance and model complexity. Codes will be available soon.