遥感影像超分算法

SRGAN

https://blog.csdn.net/qianbin3200896/article/details/104181552

SRGAN演进过程

SRCNN

ESPCN

https://blog.csdn.net/MR_kdcon/article/details/123837994

VSDR

https://blog.csdn.net/weixin_43840215/article/details/89815120

Perceptual Losses

Per-Pixel Loss: 只能提取到浅层次的特征，而无法提取到深层次的全局特征（浅层特征）
Perceptual Loss: 对进一步提取到的特征进行损失的计算（深层特征）

SRGAN

SRGAN特点：
 1. 损失函数，引入VGG特征计算MSELoss
   feature maps MSE: content details, 语义级别
  adversarial loss: texture details, 像素级别
 2. 子像素卷积放大图像
  PixelShuffle: https://blog.csdn.net/rocking_struggling/article/details/105470845

传统方法及缺陷：

• 优化目标：基于像素差异minimizing the mean squared error(MSE)
• 评价指标：peak signal-to-noise ratio(PSNR), structural similarity(SSIM)
• 缺陷：overly smooth, lack texture details, poor perceptual quality

SRGAN方法及优势：

• 优化目标：perceptual loss function = content loss + adversarial loss
• content loss：perceptual similarity instead of similarity in pixel space
• adversarial loss：differentiate between the super-resolved images and origianal photo-realistic images
• 评价：an extensive mean-opinion-score(MOS)
• 优势：recover photo-realistic

VGG网络