SA-GAN: A Second Order Attention Generator Adversarial Network with Region Aware Strategy for Real Satellite Images Super Resolution Reconstruction

Abstract

High-resolution (HR) remote sensing images have important applications in many scenarios, and improving the resolution of remote sensing images via algorithms is one of the key research fields. However, current super-resolution (SR) algorithms, which are trained on synthetic datasets, tend to have poor performance in real-world low-resolution (LR) images. Moreover, due to the inherent complexity of real-world remote sensing images, current models are prone to color distortion, blurred edges, and unrealistic artifacts. To address these issues, real-SR datasets using the Gao Fen (GF) satellite images at different spatial resolutions have been established to simulate real degradation situations; moreover, a second-order attention generator adversarial attention network (SA-GAN) model based on real-world remote sensing images is proposed to implement the SR task. In the generator network, a second-order channel attention mechanism and a region-level non-local module are used to fully utilize the a priori information in low-resolution (LR) images, as well as adopting region-aware loss to suppress artifact generation. Experiments on test data demonstrate that the model delivers good performance for quantitative metrics, and the visual quality outperforms that of previous approaches. The Frechet inception distance score (FID) and the learned perceptual image patch similarity (LPIPS) value using the proposed method are improved by 17.67% and 6.61%, respectively. Migration experiments in real scenarios also demonstrate the effectiveness and robustness of the method.

Uncertainty-driven mixture convolution and transformer network for remote sensing image super-resolution

“…Jia et al 41 designed multi-attention GAN to solve the problem that texture information of various remote sensing images is completely different. More importantly, to address the discrepancy between training data distribution and actual degraded images, Zhao et al 42 curated a genuine remote sensing dataset, enabling the training of SR models for authentic scenes. Furthermore, they introduced second-order channel attention to bolster the model’s performance in real-world scenarios.…”

SRBPSwin: Single-Image Super-Resolution for Remote Sensing Images Using a Global Residual Multi-Attention Hybrid Back-Projection Network Based on the Swin Transformer

“…Huang et al [44] proposed the Pyramid Information Distillation Attention Network (PIDAN), which employs the Pyramid Information Distillation Attention Block (PIDAB) to enable the network to perceive a wider range of hierarchical features and further improve the recovery ability of high-frequency information. Zhao et al [45] proposed the second-order attention generator adversarial attention network (SA-GAN), which leverages a second-order channel attention mechanism in the generator to fully utilize the prior information in LR images. Chen et al [46] presented the Residual Splitattention Network (RSAN), which utilizes the multipath Residual Split-attention (RSA) mechanism to fuse different channel dimensions to promote feature extraction and ensure that the network focuses more on regions with rich details.…”

A Review of GAN-Based Super-Resolution Reconstruction for Optical Remote Sensing Images

“…Recently, Zhao et al [92] presented an SR model called the second-order adversarial attention generator network (SA-GAN), which is based on real-world remote sensing imagery. The generator network of SA-GAN utilizes a second-order channel attention mechanism and a region-level nonlocal module to effectively leverage the a priori knowledge in LR images.…”