Satellite Image Super Resolution Based on Improved Deep Segmented Residual Neural Network

Authors

  • Zikang Wei, Yunqing Liu, Cong Liu

Abstract

This study employs a deep segmented residual neural network model to analyze the super-resolution of a single satellite image. A deep convolutional neural network model was analyzed, and its performance was improved. First, a Mish activation function was applied to better avoid the death and loss of neurons, which can effectively improve the generalization of neural networks. Second, use two residual layers to divide the deep network into two groups, the sum of the two residuals is the total residual, which can minimize the residual loss function and enhance the network performance. Third, optimal the learning rate was also investigated to achieve convergence more efficiently, increase the speed of iteration, and improve the overall processing speed of the model. The experimental model achieved higher peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) than networks without the proposed improvements, when tested on satellite images. Considering these results, the application of this technology will be significant for further research on satellite images.

Published

2021-06-10

Issue

Section

Articles