Edgar Salazar

Research Professor
Universidad Privada Boliviana

Google Scholar

Keynote Talk

Solving the Inverse Problem in Compressive Compton Backscattering: A Generative Diffusion Null Space Approach

Abstract

Rapid and high-fidelity body scanning is a critical requirement for security at high-throughput checkpoints. Due to their advantages in resolution, Compton scattering-based scanners are preferred when targets are continuously moving and non-static snapshots must be captured.

Recently, the sensing paradigm in Compton scanning has been revolutionized using computational imaging, through binary coded illumination. This gave rise to the Compressive X-ray Compton Backscattering Imager (CXBI), which accelerates acquisition by recovering high-quality images from a limited set of low-dimensional projections.

In this talk, we present recent advancements in CXBI, focusing on state-of-the-art Generative AI. We introduce the Invertible Diffusion Model in CXBI, which reformulates the Denoising Diffusion Null-Space Model (DDNM) as a trainable, unrolled architecture for compressed sensing, and benefits from rich hierarchical representations previously learned from extensive image datasets.

We further discuss the joint optimization of the sensing layer and diffusion-like reconstruction, and the impact of Poisson-like noise statistics on algorithmic performance. Furthermore, we share ongoing experimental results and discuss the future perspectives and challenges of deploying generative frameworks in real-world security environments.