Project Overview
Ill-posedness of ptychography is solved by
High oversampling in the data acquisition process (costly)
Exploitation of prior information about the input signal such as sparsity (not fully effective)
Hand-crafted features are shown to be limited in capability
Supervised learning-based approaches are used to exploit powerfully learned priors
Not good enough for reconstruction
Deep generative priors can capture high dimensional image distributions
Generate high dimensional samples from low dimensional latent code
Drawback: Limited representation capability
Propose an approach for x-ray ptychography leveraging the use of generative priors
Aim is to lower the overlap requirement
Further suggest adjusting the weights of generative networks to overcome the representation capability limitation of generative priors
Team