Enhanced Super-Resolution Generative Adversarial Networks https://arxiv.org/pdf/1809.00219.pdf code: https://github.com/xinntao/ESRGAN ECCV2018
在SRGAN的升级之处:
3、improved perceptual loss:VGG在激活前的feature
4、further:网络权重的插值
补充:什么是Relativistic average GAN (RaGAN):https://zhuanlan.zhihu.com/p/40403886
percpecture index是什么 和percpecture loss的关系:,percpecture是第三部分的VGG
1、Residual-in-Residual Dense Block (RDDB)
文章用了两套生成器,第一套是16 residual blocks
第二套是23 RRDB blocks
2、 Relativistic Discriminator
![](https://img-blog.csdnimg.cn/20210421214332592.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NTgyODc3MQ==,size_16,color_FFFFFF,t_70)
![](https://img-blog.csdnimg.cn/20210421214234619.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NTgyODc3MQ==,size_16,color_FFFFFF,t_70)
![](https://img-blog.csdnimg.cn/20210421215823435.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80NTgyODc3MQ==,size_16,color_FFFFFF,t_70)
Ld=
Lg=
3、improved perceptual loss
激活后太稀疏,大多数都没有被激活。
VGG采用的是ReLU激活函数,说明特征图里面很多是负的。
4、Network Interpolation
融合PSNR-oriented 的fidelity 和 GAN的perceptual quality。
α设置为0.2
注意,是网络参数的reweight,而不是两个网络最后结果的图像进行像素的reweight,实验表明 效果⑧行
code: 后期手动进行调整测试
net_PSNR = torch.load(net_PSNR_path)
net_ESRGAN = torch.load(net_ESRGAN_path)
net_interp = OrderedDict()
print('Interpolating with alpha = ', alpha)
for k, v_PSNR in net_PSNR.items():
v_ESRGAN = net_ESRGAN[k]
net_interp[k] = (1 - alpha) * v_PSNR + alpha * v_ESRGAN
torch.save(net_interp, net_interp_path)
5、训练过程:只做了x4
①cropped HR patch is 128 × 128:大图语义信息丰富
②训练PSNR-oriented model: L1loss, lr=2*10^-4, 200000 iter 减半 ?没说总共多少epoch
③训练GAN lr=1×10−4 and halved at [50k, 100k, 200k, 300k] iter