Redesigning the semi-dense residual network with multi-scale interactive for super-resolution of satellite cloud images
Bo Xia, Hao Ding, Lunqian Wang, Xinghua Wang, Weilin Liu, Zekai Zhang · 2024
High-resolution Satellite cloud images play a crucial role in weather analysis and forecast. Improving the resolution of images with super-resolution (SR) methods facilitates the weather system to identify and locate geographical information. In this paper, several SR methods have been verified on the natural color cloud images (NCCI) dataset. We propose a multi-scale interaction convolution (MIC), which realizes the interaction and adaptive aggregation of features under different reception fields. The semi-dense residual is designed to promote the dissemination of information. Based on MIC and semi-dense residual, we propose the multi-scale interaction and semi-dense residual network (MISN), which has significant advantages in both effect and parameters. The experimental results show that MISN achieves an excellent SR effect. In detail, MISN takes only 32% parameters of EDSR. The PSNR of MISN increases by up to 0.06dB and 0.5dB for ×4 SR and ×2 SR respectively compared to EDSR on the NCCI dataset. In addition, the MISN balances high accuracy and the number of parameters on the DIV2K dataset.