Object classification through scattering media with deep learning on time resolved measurement.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Traditional methods for imaging through scattering media require precise calibration of forward models, limiting their practical application.
• Variations in calibration parameters (e.g., camera field of view, illumination position) complicate and hinder accurate imaging through scattering media.

Purpose of the Study:

• To develop a calibration-invariant imaging technique for robustly identifying and classifying objects hidden behind scattering media.
• To overcome the limitations of traditional inverse problem approaches by employing a data-driven strategy.

Main Methods:

• A data-driven approach leveraging convolutional neural networks (CNNs) was employed to learn a scattering model.
• The CNN was trained on a large synthetic dataset generated using a Monte Carlo (MC) model, incorporating random variations of key calibration parameters.
• The technique was evaluated using a time-resolved camera and experimental data.

Main Results:

• The developed CNN model demonstrated invariance to calibration parameter changes within the training range and near-invariance beyond it.
• Experimental results showed successful pose estimation of a hidden mannequin, achieving 76.6% accuracy (23 out of 30 correct classifications across three poses).
• The method proved effective for non-line-of-sight (NLOS) imaging applications.

Conclusions:

• The proposed CNN-based imaging technique offers a robust and calibration-insensitive solution for non-line-of-sight imaging.
• This approach significantly advances the potential for real-time, practical NLOS imaging applications by reducing reliance on precise calibration.
• The findings pave the way for more accessible and reliable imaging systems in challenging scattering environments.