An efficient multi-scale large asymmetric-kernel network for lightweight image super-resolution

Abstract

Recently, lightweight convolutional neural networks (CNNs) on single image super-resolution (SISR) tasks have received impressive improvement with delicate structures. However, numerous lightweight methods may reduce the representation capacity of the network due to decreasing the model size and computational complexities, leading to unsatisfactory performance. In this paper, we propose an efficient multi-scale large asymmetric-kernel network (MLAN) for lightweight SISR. Specifically, MLAN is built with a succession of feature cross extraction blocks (FCEBs), which better models local and long-range interactive information of features for SR. Each of the FCEB contains a multi-scale asymmetric large-kernel attention block (MACAB) by using multiple convolutional kernels to extract features in different receptive fields and a gated mechanism to preserve the useful information for SR. Extensive experimental results on five public benchmark datasets demonstrate the superiority of MLAN over the other advanced lightweight SISR competitors. The average PSNR values are about 0.12, 0.17 and 0.11 dB greater than the second-best competitors under scaling factors of ×2, ×3 and ×4, respectively. The proposed efficient blocks enable our MLAN to make a better balance between model size and performance and achieve comparable performance with Transformer-based methods at a similar level of parameters.