Sensing Image Super-resolution Reconstruction Based on the UNet Decoding Network

To enhance the super-resolution reconstruction quality of remote sensing images, this paper fully consider the multi-scale nature of internal features and proposes a multi-scale multi-level fusion network. This network is based on the UNet encoding and decoding network, introducing multi-scale feature extraction branches to further obtain rich scale features. Simultaneously focusing on features at all levels, a fusion module is proposed using the global attention mechanism (GAM) to improve the model's feature extraction ability. The image reconstruction task is completed using sub-pixel convolution. In this paper, the UC Merced dataset is selected and reconstructed with 2x, 3x and 4x reconstruction, in which the peak signal to noise ratio (PSNR) values are 34.07 dB, 29.65 dB, and 27.63 dB, respectively, which are improved compared with some algorithms. These results further demonstrate the network presented in this paper is helpful in increasing the reconstruction quality of remote sensing images.