Real-time Terahertz Compressive Optical-Digital Neural Network Imaging

TL;DR

This paper introduces a hybrid optical-digital neural network system for real-time terahertz imaging that overcomes hardware limitations, enabling high-quality, diffraction-free, and lens-free imaging at two frames per second.

Contribution

It presents a novel THz ONN-DNN hybrid system combining physical encoding and neural networks for efficient, real-time, high-quality THz imaging with expanded field of view.

Findings

Enhanced imaging quality and diffraction-free imaging.

Real-time THz video capture at 2 fps.

Expanded field of view with lens-free design.

Abstract

Terahertz (THz) band has recently garnered significant attention due to its exceptional capabilities in non-invasive, non-destructive sensing, and imaging applications. However, current THz imaging systems encounter substantial challenges owing to hardware limitations, which result in information loss and restricted imaging throughput during data digitization and information extraction processes. To overcome these challenges, we propose a hybrid compressive optical-digital neural network designed to facilitate both real-time THz imaging and precise object information extraction. This approach utilizes a physical encoder, an optical neural network (ONN), to transform and reduce the dimensionality of physical signals, effectively compressing them to fit the physical constraints of THz sensor arrays. After the compressed signals are captured and digitized by the THz sensor array, a jointly trained digital neural network (DNN) reconstructs the signals into their desired or original form. Our proposed THz ONN-DNN computational imaging system demonstrates enhanced imaging quality, an expanded field of view with a lens-free system, diffraction-free imaging capability, and real-time THz video capture at a rate of two frames per second.