WerfNet: wide effective receptive field network for super-resolution of remote sensing images

Deep learning plays a crucial role in remote sensing image super-resolution. A network with a larger effective receptive field can extract more contextual image information, resulting in higher quality reconstructed images. Currently, stacking numerous small spatial convolutions is a popular strategy in deep learning. However, this approach only marginally widens the effective receptive field, failing to capitalize fully on the remarkable similarity of internal features in remotely sensed images. Additionally, most of the existing SR methods have the problem of underutilization of feature information. To address these problems, a residual network with wide effective receptive field for remote sensing super-resolution is proposed. Specifically, the model’s effective receptive fields are enlarged using re-parameterized large depth-wise convolution kernels to enhance contextual information extraction from remote sensing images. Feature info underutilization is overcome by fusing feature information of each level via residual connections with hierarchical structure. A high-resolution image is then reconstructed. The model’s effectiveness is validated on the aerial image dataset (AID), which showed significant improvements in visualization and objective metrics compared to several state-of-the-art methods.