CTV-Net: Complex-Valued TV-Driven Network With Nested Topology for 3-D SAR Imaging.

Area of Science:

• Electromagnetics and Remote Sensing
• Computational Imaging
• Machine Learning for Signal Processing

Background:

• Regularization methods enhance Synthetic Aperture Radar (SAR) imaging but often assume inherent sparsity, limiting accuracy for surface-like targets.
• Traditional l1 regularization struggles with precise estimations in scenarios lacking strong sparsity.
• Total Variation (TV) regularization offers edge-preserving properties beneficial for imaging.

Purpose of the Study:

• To develop a novel interpretable neural network, CTV-Net, for improved 3D SAR imaging.
• To address limitations of sparsity assumptions in existing regularization techniques.
• To enhance the precision of SAR image estimations, particularly in weakly sparse environments.

Main Methods:

• Proposed a complex-valued TV (CTV)-driven optimization model based on the 2D holography imaging operator.
• Developed a nested algorithmic framework, CTV-FIST, derived from proximal gradient descent (PGD) and FIST algorithms.
• Designed CTV-Net with layer-varied trainable weights linked to CTV-FIST hyperparameters, trained end-to-end using a cost function balancing measurement fidelity and TV norm.

Main Results:

• CTV-Net demonstrated viability and efficiency in recovering 3D SAR images from incomplete echoes.
• The method achieved precise estimations in weakly sparse scenarios, outperforming traditional approaches.
• Numerical and visual validation through extensive SAR simulations and real-measured experiments confirmed the methodology's effectiveness.

Conclusions:

• The proposed CTV-Net offers a promising regularization-based approach for high-quality 3D SAR imaging.
• CTV-Net effectively handles weakly sparse targets, overcoming limitations of standard l1 regularization.
• The interpretable, nested neural network design provides a robust framework for advanced SAR image reconstruction.