Blind Deconvolution
1
Traditional Blind Deconvolution [1990s-2015)
Image Estimation
Kernel Estimation
Conventional Alternating Minimization
Camera shake during exposure
+
=
Fronto-Parallel Scene
2
Conventional alternating minimization can be numerically unstable
Main Idea I – Something’s not right here
Image Estimation
Kernel Estimation
Iteration N
[1] Levin, Anat, et al. "Understanding blind deconvolution algorithms." IEEE TPAMI 2011
3
Main Idea II – A Toy Example
4
Main Idea III – So what should I do instead?
Fewer unknowns
5
Core Engine of our Methods
Dong, Jiangxin, et. al. "Deep wiener deconvolution: Wiener meets deep learning for image deblurring." NeurIPS 2020.
Non-Blind Solver: differentiable with respect to inputs
6
Blind Deconvolution with Diffusion Models
After each reverse diffusion step
Chung, Hyungjin, et al. "Parallel diffusion models of operator and image for blind inverse problems.“ CVPR 2023
7
Kernel Estimation Approach
8
Comparison with Blind-DPS
Metric | Blind-DPS (Alternating Estimation) | Kernel-Diff (Ours) (Kernel-First Estimation) |
PSNR 🡩 | 17.56 | 19.07 |
SSIM 🡩 | 0.387 | 0.500 |
LPIPS 🡫 | 0.583 | 0.355 |
FID 🡫 | 280.53 | 172.33 |
Total Params | 110M | 43M |
Time (in s) | 337 | 305 |
Blurred image
Blind-DPS
Kernel-Diff