Super-Resolution Reconstruction of Remote Sensing Images Based on GAN

Abstract

In recent years, the methods of super-resolution image reconstruction that based on deep learning have become a hot topic in research of computer vision. The methods of super-resolution image reconstruction that based on the Generative Adversarial Network (GAN) are not controlled in network generation, the models are easy to collapse, the generalization ability is undesirable, and the time complexity degree is too high. To fill these gaps, we propose a super-resolution image reconstruction method based on the GAN of encoding and decoding, which improves the quality of image reconstruction. First of all, our approach uses a design network with regularized structure to avoid model collapse. Then we build a generation network structure that based on encoding and decoding to suppress the uncontrollable defects of GAN network generated images. Finally, in the last layer of the generator, $\mathrm{N}^{\star}\mathrm{N}$ convolutional feature layer is included to replace the Softmax layer, which speeds up the training of the model. The experimental results show that the super-resolution remote sensing image reconstructed by the proposed method has higher reconstruction quality and better generalization ability in the DOTA training data sets. At the same time, the image reconstruction process can take much less time.