Self-Calibrating Imaging: Unsupervised Blind Deconvolution

This project tackles unsupervised blind deconvolution: jointly recovering an image and its point spread function (PSF) directly from measurements, with no clean reference. The self-calibrating structure echoes other coupled inference problems (camera poses + scene geometry in SfM, map + trajectory in SLAM, spike times + waveforms in spike sorting) where the unknown of interest can only be disentangled alongside a calibration parameter.

The forward model is y = k * x + n: same-padding 2D convolution with the PSF plus optional Gaussian noise. A MAP objective is optimized jointly over the image and kernel, with kernel priors (L2, center-of-mass, autocorrelation), a pink-noise image prior, a custom image-prior hook, and an optional DDPM diffusion prior. Randomized optics (PSFs whose autocorrelation is close to a delta) help disambiguate scene content even when the exact PSF realization is unknown.

PSF generators cover Gaussian, linear motion (length and angle), atmospheric turbulence (Fried parameter with distortion seeds), and randomized optics (band-limited Fourier phases), with an identity-blur option for sanity checks. A testbench sweeps PSF types, kernel sizes, learning rates, and prior weights across every image in the dataset, logging PSNR, SSIM, kernel error, and reconstruction artifacts to Weights & Biases. Slurm orchestrates the sweep across GPU partitions, keeping the first job that starts and dropping the rest.