Passive non-line-of-sight imaging with diffuse-aware attention-enhanced encoding

Area of Science:

• Computer Vision
• Signal Processing
• Optics

Background:

• Passive non-line-of-sight (NLOS) imaging reconstructs targets hidden from direct view using ambient light.
• Traditional methods struggle with low signal-to-noise ratio (SNR) and signal attenuation in deep learning models.

Purpose of the Study:

• To develop a novel attention mechanism for passive NLOS imaging that addresses limitations of generic attention modules.
• To improve the quality of reconstructed images from occluded scenes by incorporating physical principles.

Main Methods:

• Proposed a diffuse-aware attention module (DAAM) integrating physical priors of diffuse reflections.
• Utilized deformable convolution for anisotropic spatial attention and mean-std pooling for frequency-aware channel attention.
• Integrated DAAM into a residual-attention encoder architecture.

Main Results:

• DAAM significantly improved performance compared to traditional passive NLOS methods and generic attention mechanisms.
• Achieved higher peak signal-to-noise ratio (PSNR) and better perceptual fidelity (LPIPS).
• Demonstrated effective preservation of weak signals and enhancement of discriminative features.

Conclusions:

• Incorporating physical priors into attention module design is crucial for high-quality passive NLOS reconstruction.
• DAAM offers a promising approach for enhancing passive NLOS imaging capabilities.
• The method shows potential for applications requiring reconstruction of occluded scenes.