Multi-Receptive-Fields Convolutional Network for Remote Sensing Images Super-Resolution

Recently, single image super-resolution (SISR) has been widely applied in the field of remote sensing image processing and obtained remarkable performance, focusing on restoring the high-resolution (HR) image from a low-resolution (LR) image. However, we observe that the existing CNN-based SISR methods mainly focus on wider or deeper architecture design, neglecting to exploit features at global receptive field. Moreover, the LR inputs and features contain abundant low-frequency information, which are perceived equally in the same receptive field, hence limiting the representational ability of CNNs. To solve these problems, we propose a Multi-Receptive-Fields Super Resolution Network (MRFSR) for remote sensing image reconstruction. The proposed network employs non-local neural network to enhance low-level complex features by expanding the receptive field of the shallow convolution layer. Moreover, we propose the multi-branch up- and down-sampling modules to deal with LR features in multiple receptive fields, which can enhance the high-frequency components and learn abstract feature representations in multiple scales, respectively. Extensive experiments on NPU-RESISC45 dataset shows that the proposed MRFSR can provide state-of-the-art or even better performance in both quantitative and qualitative measurements.