Super-Resolution图像/视频超分

AI论文简报2026年6月7日2

Super-Resolution图像/视频超分

数据预处理

降质方法(Degradation)

  • Residual Dense Network for Image Super-Resolution
    • resize by bicubic
    • Gaussian Blur kernel 7x7
    • Gaussian Noise level 30
  • Old Photo Restoration via Deep Latent Space Translation
    1. Gaussian white noise with (5, 50)
    2. Gaussian blur with kernel size (3, 5, 7) and standard deviation (1,0, 5.0)
    3. JPEG compression whose quality level in the range of (40; 100);
    4. Color jitter which randomly shifts the RGB color channels by (-20; 20);
    5. Box blur to mimic the lens defocus

数据增强(Data augmentation)

  • flip: flipping horizontally or vertically

  • rotating旋转: 90 180 270

  • batch-size: 16

  • image size: 48x48 / 32x32

  • adam: init lr: 10-4, after 200 epoch: 5x10-5

  • mini-batch size of SGD to 64, momentum parameter to 0.9, and weight decay 10-4

评价指标

PSNR

SSIM

LPIPS

FID

数据集

  • training set: DIV2K dataset

  • standard benchmark testing datasets

    • Set5
    • Set14
    • B100
    • Urban100
    • Manga109
  • https://aijishu.com/a/1060000000164177

  • Learning deep cnn denoiser prior for image restoration

  • Spatial pyramid pooling in deep convolutional networks for visual recognition

  • L. Zhu, “pytorch-opcounter,” https://github.com/Lyken17/ pytorch-OpCounter, 2019.

经典算法

EDSR:Enhanced Deep Residual Networks for Single Image Super-Resolution

RDN:Residual Dense Network for Image Super-Resolution

意图

Residual learning + Densnet

网络结构设计

全局残差学习

所有RDB的全局fusion(GFF)

  • RDB: 20个
RDB: residual Block + dense Block

  • RDB中conv: 6个
  • Densnet Grow rate: 32
不同结构的比较

CM的只是dense connection(Densnet)

训练

  • flip: flipping horizontally or vertically
  • rotating旋转: 90 180 270
  • batch-size: 16
  • image size: 48x48 / 32x32
  • adam: init lr: 10-4, after 200 epoch: 5x10-5

结果

RCAN:Image Super-Resolution Using Very Deep Residual Channel Attention Networks

意图

EDSR通过构建深度的残差模块获得显著的SR性能的提升,意图引入有效利用channel-wise features来构建有效的深度SR模型。

  • high-frequency channel-wise features are more informative for HR reconstruction.
  • It also indicates that deeper networks may be easier to achieve better performance than wider networks.

网络结构设计

CA channel attention

LR中有冗余的低频信息和有价值的高频信息,如边缘,纹理和其他细节。 利用attention来甄别信息权重。 CA基本和SE-Net的注意力机制一样。

RCAB = CA + Residual Block

总体结构:全局残差和局部残差

  • RG数量: 10
  • RCAB: 20
  • number of feature map: 64

训练

结果

SRGAN + SRResNet:Photo-realistic single image superresolution using a generative adversarial network

意图

  • how do we recover the finer texture details when we super-resolve at large upscaling factors?

  • 亮点

    • perceptual loss function
    • GAN

网络结构设计

从上可知网络分为生成网络(SRRestnet) 和判别网络, 各个参数在图中有清晰的标明。生成网络(SRRestnet)采用的是残差模式,残差模块B=16

perceptual loss function

最小化基于MSE的loss, 鼓励像素的均值化,导致平滑模糊,缺少感知质量。

perceptual loss function 包含内容loss和对抗loss

  • 内容loss MSE loss

    内容VGG loss:VGG 19 特征层的MSE loss

  • 对抗loss 对抗loss就是正常的对抗网络的loss

训练

  • 训练数据:350 thousand images from ImageNet
  • batch size : 16
  • input size: HR(9696) LR(2424)
  • Adam: 0.9
  • SRResnet: lr 1e-4, 1e6 更新一次 1e-5
  • loss:选择不同的loss(mse,感知loss(vgg 不同层vgg22, vgg54)

结果

PSNR不能是唯一标准,PSNR高画面质量人为感知不一定好。