Step1 分别训练
V
A
E
1
{VAE}_1
VAE1与
V
A
E
2
{VAE}_2
VAE2
目标是学习R到Y的映射,即给定一个真实老照片r,得到清晰的修复结果。
为了解决合成照片与真实照片域间差异导致的泛化性问题,X与R共享同一个VAE1(图中绿色部分domain gap标注有误,应标注在非交集的部分)。其中,VAE比普通自编码器更容易获得密集分布的隐空间表示,这样
Z
X
Z_X
ZX与
Z
R
Z_R
ZR之间的domain gap会更小。训练一个鉴别器来区分
Z
X
Z_X
ZX与
Z
R
Z_R
ZR,以进一步缩小domain gap。
总目标函数为
Step2 固定
V
A
E
1
{VAE}_1
VAE1与
V
A
E
2
{VAE}_2
VAE2,独立训练隐空间映射T
目的是为了更容易缩小domain gap。其中从左往右逐渐增加组件。1)Pix2Pix:translation in image-level 2)two VAEs with an additional KL loss to penalize the latent space 3)两阶段训练,VAEs权重不固定 4)完整模型,采用了latent adversarial loss。
Strong conclusions: We propose a novel triplet domain translation network to restore the mixed degradation in old photos. The domain gap is reduced between old photos and synthetic images, and the translation to clean images is learned in latent space. Our method suffers less from generalization issue compared with prior methods.
三个域之间的图像转换网络
缩小隐空间domain gap,并在隐空间学习到更佳的图像转换结果
泛化性更强
Weak conclusions: We propose a partial nonlocal block which restores the latent features by leveraging the global context, so the scratches can be inpainted with better structural consistency.
M. Suganuma, X. Liu, and T. Okatani, “Attention-based adaptive selection of operations for image restoration in the presence of unknown combined distortions,” arXiv preprint arXiv:1812.00733, 2018.
D. Ulyanov, A. Vedaldi, and V. Lempitsky, “Deep image prior,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 9446–9454.
T.-C. Wang, M.-Y. Liu, J.-Y. Zhu, A. Tao, J. Kautz, and B. Catanzaro, “High-resolution image synthesis and semantic manipulation with conditional gans,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 8798–8807.
J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 2223–2232.
K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Bm3d image denoising with shape-adaptive principal component analysis,” 2009.
K. Nazeri, E. Ng, T. Joseph, F. Qureshi, and M. Ebrahimi, “Edgeconnect: Generative image inpainting with adversarial edge learning,” 2019.
Attention-based Adaptive Selection of Operations for Image Restoration in the Presence of Unknown Combined Distortions(CVPR2019)
image-to-image translation图像转换
Deal with multiple combined distortions with unknown mixture ratios
并行执行多个操作,这些操作由注意力机制加权,根据输入选择适当的操作。
Deep Image Prior(CVPR2018)
Standard inverse problems such as denoising, super-resolution, and inpainting。一系列image restoration问题都是Standard inverse problems。
由于网络学信号学得快,学噪声学得慢,所以要利用这一点控制好迭代次数:不能太少,否则学不到位,输出结果结构模糊;也不能太多,否则学的太好,跟
x
0
x_0
x0过于相似,把噪声也一起学了。其中
x
0
x_0
x0为corrupted image。
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs(CVPR2018)
image-to-image translation, whose goal is to translate an input image from one domain to another domain given input-output image pairs as training data
pix2pix framework
coarse-to-fine generator
multi-scale discriminator architecture, and a robust adversarial learning objective function。判别器有三个,针对不同尺寸图像。三个尺寸分别是原尺寸,二分之一,四分之一。receptive field大小的问题。