Large coordinate attention network for lightweight image super-resolution

The Multi-Scale Receptive Field (MSRF) and Large Kernel Attention (LKA) module have been shown to significantly improve performance in the lightweight image super-resolution (SR) task. However, existing lightweight SR methods seldom pay attention to designing lightweight yet effective building block with MSRF for local modeling, and their LKA modules face a quadratic increase in computational and memory footprints as the convolutional kernel size increases. To address the first issue, we propose a simple but effective block, Multi-scale Blueprint Separable Convolutions (MBSConv), as highly efficient building block with MSRF, and it can focus on the learning for the multi-scale information which is a vital component of discriminative representation. As for the second issue, in order to mitigate the complexity of LKA, we propose a Large Coordinate Kernel Attention (LCKA) module which decomposes the two-dimensional convolutional kernels of the depth-wise convolutional layers in LKA into horizontal and vertical one-dimensional kernels. LCKA enables the adjacent direct interaction of local information and long-distance dependencies not only in the horizontal direction but also in the vertical. Besides, LCKA allows for the direct use of extremely large kernels in the depth-wise convolutional layers to capture more contextual information which helps to significantly improve the reconstruction performance, while incurring lower computational complexity and memory footprints. Integrating MBSConv and LCKA, we propose a Large Coordinate Attention Network (LCAN) which is an extremely lightweight SR network with efficient learning capability for local, multi-scale, and contextual information. Extensive experiments show that our LCAN with low model complexity achieves superior performance compared to previous lightweight state-of-the-art SR methods.