Double Blind Imaging with Generative Modeling

TL;DR

This work tackles blind inverse imaging where the forward operator $A$ is unknown by learning a probabilistic model $p(A)$ from unpaired clean images and corrupted measurements using an AmbientGAN-inspired framework. A generator produces forward operators, forming synthetic measurements that are indistinguishable from real measurements, effectively yielding a physics-informed prior over degradation processes. The learned forward-operator prior is then used as a plug-in in diffusion-based blind deconvolution, enabling accurate reconstructions even without paired training data. Experiments on Gaussian and motion blur demonstrate robust PSF identification and competitive downstream recovery under various noise levels, highlighting the method's modularity and potential applicability to broader structured blind inverse problems.

Abstract

Blind inverse problems in imaging arise from uncertainties in the system used to collect (noisy) measurements of images. Recovering clean images from these measurements typically requires identifying the imaging system, either implicitly or explicitly. A common solution leverages generative models as priors for both the images and the imaging system parameters (e.g., a class of point spread functions). To learn these priors in a straightforward manner requires access to a dataset of clean images as well as samples of the imaging system. We propose an AmbientGAN-based generative technique to identify the distribution of parameters in unknown imaging systems, using only unpaired clean images and corrupted measurements. This learned distribution can then be used in model-based recovery algorithms to solve blind inverse problems such as blind deconvolution. We successfully demonstrate our technique for learning Gaussian blur and motion blur priors from noisy measurements and show their utility in solving blind deconvolution with diffusion posterior sampling.

Figures (9)

• Figure 1: Example of imaging system setup where the measurement process is unknown due to hardware, environment, or scene variables.
• Figure 2: Schematic block diagram illustrating our algorithm for unsupervised learning of the imaging system parameters. We learn a generative network $G_\theta$ that generates system parameters $\kappa^g$. Given a set of training images $\{x_1, x_2, \cdots, x_N\}$, we pass these images along with the parameters $\kappa^g$ through the imaging system to generate measurements $y^g$. The discriminative network $D_\phi$ is trained to distinguish between the generated measurements $y^g$ and actual measurements $y^r \sim\{y_1, y_2, \cdots, y_N\}$. Note that the image dataset and measurement dataset are unpaired, independent, and disjoint of one another.
• Figure 3: BDPS reconstructions for blind Gaussian (top) and motion (bottom) deblurring on AFHQ $128\times 128$. Using various kernel priors.
• Figure 4: Ground truth point spread functions for both the Gaussian (left) and motion (right) examples.
• Figure 5: Example training split of unpaired data where the top row is for Gaussian blur and the bottom row is motion blur.