Hyperspectral Image Super-Resolution for Multi Attention-Generative Adversarial Network

Image super-resolution (SR) methods can generating a remote sensing imaging from high spatial resolution without increasing a cost, these are difficult to obtain due to high cost of accession equipment and complex weather. In this paper, this model proposes a Multi Attention-Generative Adversarial Network (MA-GAN) on cave dataset were used. The proposed model first designed of network framework for image super-resolution task. If generator contain a two blocks: Pyramidal Convolution in Residual-Dense Block (PCRDB) and Attention-Based Up Sample Block (AUP). The PCRDB module combining a multi-scale convolution and channel attention to automatically learning and adjusting a scaling of the residual for better result. Finally, this model present a loss function based on pixel loss and presented a both adversarial loss and feature loss to guide the generating learning. The result obtained MA-GAN show a PSNR was improved of 46.10, RMSE of 1.77, SAM of 5.21, ERGAS of 0.15 and SSIM of 2.14 on cave dataset, which ensure GAN compared to other existing methods like Two-Stream Function Network (TSFN), Separable-Spectral and Inception Network (SSIN) and Non-Negative Clustering Sparse Representation (NNCSR).