Super-resolution of SAR images using residual four-channel large kernel attention

The advantages of Synthetic Aperture Radar (SAR) include its ability to produce high-resolution images in all-weather conditions and its wide range of applications in both military and civilian fields. However, SAR images are frequently subject to degradation in resolution due to the presence of speckle noise. This paper presents a new super-resolution network, the residual four-channel large kernel attention (RFCLKA) network, which uses four different-scale large kernel attentions (LKA) as primary components and integrates residual learning. The RFCLKA network can learn feature information at multiple scales, allowing the network to focus on informative features and improve their representation. Moreover, residual learning in RFCLKA supports deeper network construction and efficient network optimization, enabling effective learning of global and local features. The proposed SAR super-resolution network further incorporates the bilinear interpolation results of low-resolution SAR images as auxiliary information to provide base features, accelerating training speed and improving network performance. Through ablation experiments and comparative experiments, the proposed network’s effectiveness is demonstrated based on objective metrics and visual quality of the super-resolved SAR images.