当前位置: X-MOL 学术Int. J. Aeronaut. Space Sci. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Satellite Imagery Super-Resolution Using Squeeze-and-Excitation-Based GAN
International Journal of Aeronautical and Space Sciences ( IF 1.4 ) Pub Date : 2021-06-23 , DOI: 10.1007/s42405-021-00396-6
Marwa S. Moustafa , Sayed A. Sayed

Single Image Super Resolution (SISR) elevates spectral and spatial image resolution beyond the sensor capabilities. Convolutional Neural Networks (CNNs) have dominated current mainstream approaches. However, the utilization of pixel-based loss function hinders achieving realistic perceptual results at large upscale factors. Recently, Generative Adversarial Network (GAN) attained more realistic, crisp results in natural image, but the complex nature of satellite images limits the performance. In this work, we address these challenges equipped with the promising results of Squeeze-and-Excitation (SE) in classification tasks. A Spatial and Channel Squeeze-and-Excitation GAN (SCSE-GAN) is introduced. The proposed generator stacked SCSE block after each residual block to recalibrate and ensure features flow and amplify high-frequency details. In addition, skip/residual connection was utilized in the GAN generator network to further boost the performance. Wasserstein distance with gradient penalty (WGAN-GP) was adopted to stabilize training and avoid gradient vanishing phenomena. Finally, we conducted various experiments on two open-source benchmarks namely: RSSCN7 and Kaggle datasets, to systematically evaluate the proposed framework performance. According to the obtained results, the proposed approach excels other approaches quantitatively and visually. Quantitatively, the results show a boost by a considerable margin of 2%, and 3% in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), respectively. Visually, the proposed method shows a sharper, less smooth image compared with benchmark SISR approaches.



中文翻译:

使用基于挤压和激励的 GAN 的卫星图像超分辨率

单图像超分辨率 (SISR) 将光谱和空间图像分辨率提升到传感器能力之外。卷积神经网络 (CNN) 主导了当前的主流方法。然而,基于像素的损失函数的利用阻碍了在大的高档因素下实现真实的感知结果。最近,生成对抗网络(GAN)在自然图像中获得了更逼真、更清晰的结果,但卫星图像的复杂性限制了性能。在这项工作中,我们解决了这些挑战,并配备了在分类任务中挤压和激发 (SE) 的有希望的结果。引入了空间和通道挤压和激励 GAN (SCSE-GAN)。建议的生成器在每个残差块之后堆叠 SCSE 块,以重新校准并确保特征流动并放大高频细节。此外,在 GAN 生成器网络中使用了跳过/残差连接以进一步提高性能。采用梯度惩罚的 Wasserstein 距离(WGAN-GP)来稳定训练并避免梯度消失现象。最后,我们对两个开源基准进行了各种实验,即:RSSCN7 和 Kaggle 数据集,以系统地评估所提出的框架性能。根据获得的结果,所提出的方法在数量和视觉上都优于其他方法。从数量上讲,结果显示峰值信噪比 (PSNR) 和结构相似性指数度量 (SSIM) 分别提高了 2% 和 3%。在视觉上,与基准 SISR 方法相比,所提出的方法显示出更清晰、更不平滑的图像。GAN 生成器网络中使用了跳过/剩余连接以进一步提高性能。采用梯度惩罚的 Wasserstein 距离(WGAN-GP)来稳定训练并避免梯度消失现象。最后,我们对两个开源基准进行了各种实验,即:RSSCN7 和 Kaggle 数据集,以系统地评估所提出的框架性能。根据获得的结果,所提出的方法在数量和视觉上都优于其他方法。从数量上讲,结果显示峰值信噪比 (PSNR) 和结构相似性指数测量 (SSIM) 分别提高了 2% 和 3%。在视觉上,与基准 SISR 方法相比,所提出的方法显示出更清晰、更不平滑的图像。GAN 生成器网络中使用了跳过/剩余连接以进一步提高性能。采用梯度惩罚的 Wasserstein 距离(WGAN-GP)来稳定训练并避免梯度消失现象。最后,我们对两个开源基准进行了各种实验,即:RSSCN7 和 Kaggle 数据集,以系统地评估所提出的框架性能。根据获得的结果,所提出的方法在数量和视觉上都优于其他方法。从数量上讲,结果显示峰值信噪比 (PSNR) 和结构相似性指数测量 (SSIM) 分别提高了 2% 和 3%。在视觉上,与基准 SISR 方法相比,所提出的方法显示出更清晰、更不平滑的图像。

更新日期:2021-06-23
down
wechat
bug