Bio-Inspired Ghost Imaging: A Self-Attention Approach for Scattering-Robust Remote Sensing.

Area of Science:

• Computational imaging
• Deep learning for remote sensing
• Bio-inspired algorithms

Background:

• Ghost imaging (GI) is effective for remote sensing but struggles with atmospheric scattering (fog).
• Fog causes noise and signal attenuation, limiting traditional GI performance.
• Biological visual systems use selective attention to filter noise and focus on relevant information.

Purpose of the Study:

• To develop a novel deep learning architecture for ghost imaging in foggy environments.
• To incorporate a self-attention mechanism inspired by biological visual systems.
• To enhance image reconstruction quality and robustness under scattering conditions.

Main Methods:

• A deep learning model with an embedded self-attention mechanism was designed.
• The model processes one-dimensional bucket measurements to capture local and global dependencies.
• Simulations were conducted on MNIST and a Human-Horse dataset under varying fog densities and sampling rates.

Main Results:

• The bio-inspired model outperformed conventional GI and CNN-based methods in foggy conditions.
• Achieved Peak Signal-to-Noise Ratio (PSNR) of 24.5-25.5 dB/m and Structural Similarity Index Measure (SSIM) of ~0.8 at β ≥ 3.0 dB/m and N ≥ 50%.
• Demonstrated computational efficiency with inference times under 0.12 seconds.

Conclusions:

• The self-attention module is crucial for high-quality ghost imaging reconstruction in adverse weather.
• The proposed bio-inspired deep learning approach sets a new benchmark for computational imaging.
• Potential applications include environmental monitoring, autonomous navigation, and defense systems.