Non-line-of-sight imaging based on an untrained deep decoder network

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Non-line-of-sight (NLOS) imaging aims to visualize objects or scenes not directly visible to a sensor.
• Existing NLOS methods often require complex setups, extensive training data, or are limited by ambient light conditions.
• Low-cost, high-quality passive NLOS imaging remains a significant research challenge.

Purpose of the Study:

• To develop a novel reconstruction method for occluder-aided NLOS imaging.
• To enable high-quality scene reconstruction using an untrained deep decoder network.
• To overcome limitations of existing NLOS techniques, particularly in high ambient light and without training data.

Main Methods:

• Utilized an untrained deep decoder network integrated with the physical forward model of NLOS imaging.
• Employed a method where network weights are automatically updated through interaction with the forward model, eliminating the need for training datasets.
• Applied the technique to reconstruct hidden scenes from photographs taken under high ambient light conditions.

Main Results:

• Achieved high-quality reconstructions of hidden scenes.
• Demonstrated superior performance in terms of image detail and robustness compared to existing methods through simulations and experiments.
• Successfully reconstructed scenes from images of a blank wall, showcasing the method's effectiveness in challenging conditions.

Conclusions:

• The proposed untrained deep decoder network offers a powerful and data-efficient approach to NLOS imaging.
• The method shows significant promise for practical applications of NLOS imaging in real-world scenarios.
• This advancement contributes to the development of low-cost, high-quality passive NLOS imaging systems.