Inverse Problems in Imaging under Pretrained Generative Prior

Post by: admin
December 27, 2021
Comments off

Imaging inverse problems arise in a myriad of applications in signal processing, and computer vision. In particular, image dehazing, and deblurring is required in astronomy to recover sharp images from blurred, hazy, observed galaxy images; image retrieval and denoising from partial measurements for MRI or CT scan in medical imaging; image super resolution, and inpainting for human face recognition from blurred out, low resolution, and corrupted images. The main idea is to restore an image from its partial, incomplete, and noisy observations. In the absence of additional information image restoration could be very challenging or even impossible, for example, it might be impossible to recover the registration number of a fast moving vehicle from a blurred snapshot of its number plate. We employ neural networks to assist classical image recovery algorithms to recover very good estimates of the true images from incomplete or partial corrupted measurements.

In particular, deep generative models, especially trained to learn image distributions assist in the range of image restoration tasks discussed above. As generative models, we have studied generative adversarial networks, and ow based invertible neural networks as pretrained prior in imaging inverse problem and have obtained state-of-the art image restoration results.

This project involves techniques and tools from classical signal processing, and more recent machine/deep learning. We work on developing actual working models using Python, Tensorow, Pytorch, Matlab, and also carefully study the technical aspects of these problems using tools from statistics, optimization theory, signal processing, and computational science, in general.

Applications

Surveillance, Optical Imaging, X-ray crystallography, Long-distance Imaging.

[2] M. Asim, F. Shamshad, and A. Ahmed Blind image deconvolution using deep generative priors, arXiv preprint arXiv:1802.04073, 2018.

[3] M. Asim, A. Ahmed, and P. Hand Invertible generative models for inverse problems: mitigating representation error and dataset bias, arXiv preprint arXiv:1905.11672, 2019.

[4] F. Shamshad, and A. Ahmed Robust Compressive Phase Retrieval via Deep Generative Priors, submitted, 2019.

[5] F. Shamshad, F. Abbas and A. Ahmed Deep Ptych: Subsampled Fourier Ptychography via Generative Priors, In International Conference on Acoustic, Speech, and Signal Processing (ICASSP), 2019.

[6] F. Shamshad, A. Hanif, F. Abbas, M. Awais, and A. Ahmed Image Adaptive Generative Priors for Subsampled Fourier Ptychography., Full-length paper accepted in proceedings of International Conference on Computer Vision (ICCV) workshop on Learning for Computational Imaging (LCI), 2019.

[7] F. Shamshad, and A. Ahmed Phase Retrieval through Scattering Media via Generative Models. Poster accepted in Computer Vision and Pattern Recognition (CVPR) workshop on Computational Cameras and Displays (CCD), arXiv preprint arXiv:1802.04073, 2019.

[8] F. Shamshad, F. Abbas, and A. Ahmed Subsampled Fourier Ptychography Using Generative Priors, Poster accepted in Computer Vision and Pattern Recognition (CVPR) workshop on Computational Cameras and Displays (CCD), 2019.