Remote Sensing Image Super-Resolution Adversarial Network Based on Reverse Feature Fusion and Residual Feature Dilation

Enhancing the resolution of images by super-resolution reconstruction algorithm is less costly compared with the way of upgrading hardware devices. At present, the image super-resolution algorithm can recover the texture details of the image well, but the performance is relatively general on the remote sensing images with lower contrast and more complex texture, and the reconstructed images are prone to noise, missing texture and checkerboard effect. This paper proposes a super-resolution adversarial network based on inverse feature fusion for the characteristics of remote sensing images, which combines the high-level semantics of the image with the low-level semantics to make the reconstructed images perform better in overall performance, and enhance the local details of the reconstructed image by feature expansion module. Experiments are carried out on RSSCN7 Dataset and representative algorithms are used for comparison. Experimental results show that RFSRNet and RFSRGAN offer better results than other methods, and ablation study show proposed modules are effective in enhancing the perceived similarity.