Reconstruction of super-resolution from high-resolution remote sensing images based on convolutional neural networks

In this study, a novel algorithm named the Edge-enhanced Generative Adversarial Network (EGAN) is proposed to address the issues of noise corruption and edge fuzziness in the super-resolution of remote sensing images. To build upon the baseline model called Deep Blind Super-Resolution GAN (DBSR-GAN), an edge enhancement module is introduced to enhance the edge information of the images. To enlarge the receptive field of the algorithm, the Mask branch within the edge enhancement structure is further optimized. Moreover, the loss of image consistency is introduced to guide edge reconstruction, and subpixel convolution is employed for upsampling, thus resulting in sharper edge contours and more consistent stylized results. To tackle the low utilization of global information and the reconstruction of super-resolution artifacts in remote sensing images, an alternative algorithm named Nonlocal Module and Artifact Discrimination EGAN (END-GAN) is proposed. The END-GAN introduces a nonlocal module based on the EGAN in the feature extraction stage of the algorithm, enabling better utilization of the internal correlations of remote sensing images and enhancing the algorithm's capability to extract global target features. Additionally, a method discriminating artifacts is implemented to distinguish between artifacts and reals in reconstructed images. Then, the algorithm is optimized by introducing an artifact loss discrimination alongside the original loss function. Experimental comparisons on two datasets of remote sensing images, NWPUVHR-10 and UCAS-AOD, demonstrate significant improvements in the evaluation indexes when the proposed algorithm is under investigation.