A Physics-Informed Deep Learning Approach for Real-Time Aerial Defense
This project addresses a critical challenge in automated defense: spotting and pinpointing ballistic missiles in real-time via optical sensors. Missiles appear as dim, short-lived lines on Focal Plane Arrays, often obscured by heavy distortions from air interference, Poisson sensor noise, and background clutter.
Conventional techniques like Maximum Likelihood Estimation (MLE) or GLRT fall short due to high computational demands and performance drops under weak signal-to-noise ratios (SNR). To overcome this, we propose SAR-Net (Spatial Attention Residual Network)—a system that learns patterns directly from data, enabling fast and dependable solutions where speed is vital.
- Learning from Weak Signals: Traditional modeling falters under noise. Our goal is teaching a Neural Network to distinguish streaks from clutter using learned connections rather than rigid theory-bound assumptions.
- Trajectory Estimation: Predicting curved flight paths shaped by air resistance requires high flexibility, which simple straight-line methods lack.
- Unified Model: Moves beyond sequential autoencoder-detector systems to a single-step inference model.
- Edge Deployment: Cutting response times to milliseconds, making compact AI models viable for hardware like C-RAM (Counter Rocket, Artillery, and Mortar).
SAR-Net integrates a unique Spatial Attention Module (SAM) into residual blocks to prioritize meaningful features over background interference.
- Residual Feature Extraction: Built using ResBlocks to enable deeper layers without losing signal strength.
- Spatial Attention Mechanism: Unlike standard CNNs, SAM identifies the "streak" and downweights irrelevant noise before making decisions.
- Dual-Head Prediction:
- Classification Head: Generates likelihood via Sigmoid activation.
- Regression Head: Forecasts smooth position vectors for the trajectory.
Note
**Architecture Diagram **
Figure 1: Proposed SAR-Net architecture featuring Residual Blocks and SAM.
Since actual missile data is restricted, we utilize a simulator grounded in physical laws:
- Non-Linear Trajectory: Bezier Curves capture air resistance and gravity effects.
- Realistic Noise: Gaussian PSFs for the main signal and Poisson layers for atmospheric scattering.
- Multi-Objective Loss: Minimizes Binary Cross-Entropy (Detection) and Mean Squared Error (Localization).
- Optimization: Adam optimizer with weighted loss (Classification 1.0, Regression 2.0).
SAR-Net is measured side-by-side with classic methods across 1,000 diverse test instances.
- ROC Curves: Demonstrates detection accuracy under varying noise conditions (2 dB to 15 dB).
- Visual Verification: Overlaying red predictions on green ground truth to verify precision.
| Feature | Conventional (GLRT/MLE) | Proposed SAR-Net |
|---|---|---|
| Path Modeling | Straight-line only | Non-linear Bezier Curves |
| Speed | Slow (Sequential scans) | Real-time (Milliseconds) |
| Low SNR | High failure rate | Attention-driven robustness |
- Sim-to-Real Gap: Addressed through heavy augmentation and Poisson-Gaussian noise blends.
- Complexity: Navigating the loss surface for non-linear manifolds.
- Hardware Efficiency: Optimized for NVIDIA T4/RTX GPUs to maintain <10ms latency.
Important
Missile Streak Detection and Localization System U.docx
Section Highlights:
- 2.4 Regression Complexity: Details on navigating non-linear loss surfaces.
- 3.7 Ablation Study: Proof of SAM module contribution in reducing interference.
- 3.8 Monte Carlo Simulation: Statistical validation strategy.
- [1] Balci & Tekalp, "Machine learning for missile streak detection," ICIP, 2021.
- [3] Zhang et al., "Attention-guided pyramid context networks," IEEE TAES, 2021.
- [6] K. He et al., "Deep residual learning for image recognition" (ResNet), CVPR, 2016.