DLKN:Enhanced Lightweight Image Super-Resolution with Dynamic Large Kernel Network

Abstract

Convolutional Neural Networks (CNNs) are constrained in adaptively capturing information due to the use of fixed-size kernels. Although they provide a wide receptive field and achieve competitive performance with fewer parameters by using decomposed large kernels, they lack adaptability. Therefore, we propose the Dynamic Large Kernel Network (DLKN) for lightweight image super-resolution. Specifically, we design a basic convolutional block of feature aggregation groups, akin to the transformer architecture. It comprises a dynamic large kernel attention block and a local feature enhancement block that can adaptively utilize information. In our dynamic large kernel attention block, we decompose the large kernel convolution into kernels with different sizes and expansion rates. We then fuse their information for weight selection, dynamically adjusting the proportion of information from different receptive fields. The local feature enhancement block significantly improves local feature extraction with low parameter counts. It encourages interactions between local spatial features by decomposing the convolution into horizontally and vertically cascading kernels. Experimental results on benchmark datasets demonstrate that our proposed model achieves excellent performance in lightweight and performance-oriented super-resolution tasks. It successfully balances the relationship between performance and model complexity. The code is available at https://github.com/LyTinGiu/DLKN_SR.