Knowledge-Informed Neural Network for Complex-Valued SAR Image Recognition

TL;DR

The Knowledge-Informed Neural Network (KINN) is introduced, a lightweight framework built upon a novel compression-aggregation-compression architecture that establishes a state-of-the-art in parameter-efficient recognition, offering exceptional generalization in data-scarce and out-of-distribution scenarios and tangible interpretability.

Abstract

Deep learning models for complex-valued Synthetic Aperture Radar (CV-SAR) image recognition are fundamentally constrained by a representation trilemma under data-limited and domain-shift scenarios: the concurrent, yet conflicting, optimization of generalization, interpretability, and efficiency. Our work is motivated by the premise that the rich electromagnetic scattering features inherent in CV-SAR data hold the key to resolving this trilemma, yet they are insufficiently harnessed by conventional data-driven models. To this end, we introduce the Knowledge-Informed Neural Network (KINN), a lightweight framework built upon a novel "compression-aggregation-compression" architecture. The first stage performs a physics-guided compression, wherein a novel dictionary processor adaptively embeds physical priors, enabling a compact unfolding network to efficiently extract sparse, physically-grounded signatures. A subsequent aggregation module enriches these representations, followed by a final semantic compression stage that utilizes a compact classification head with self-distillation to learn maximally task-relevant and discriminative embeddings. We instantiate KINN in both CNN (0.7M) and Vision Transformer (0.95M) variants. Extensive evaluations on five SAR benchmarks confirm that KINN establishes a state-of-the-art in parameter-efficient recognition, offering exceptional generalization in data-scarce and out-of-distribution scenarios and tangible interpretability, thereby providing an effective solution to the representation trilemma and offering a new path for trustworthy AI in SAR image analysis.

Figures (14)

• Figure 1: Comparison of KINN with state-of-the-art methods, including lightweight CNNs (green), ViTs (orange), and hybrid models for SAR (blue). KINN achieves a better performance-efficiency trade-off. All models are evaluated on the MSTAR dataset under the challenging OFA-3 protocol, using only 10% of the available training data.
• Figure 2: A visualization of Equation \ref{['equ:sparse_rep']} in both the frequency and time domains. The upper row illustrates the frequency domain components, while the lower row depicts the image domain components subsequent to the use of Chirp Scaling (CS) algorithm and Inverse Fast Fourier Transform (IFFT).
• Figure 3: An overview of the proposed KINN architecture. The framework (a) follows a compression-aggregation-compression paradigm. The initial compression is achieved by (b) an ISTA-based unfolding network that incorporates physical priors. Following aggregation, features are further compressed in a latent space, where each self-distillation branch composed of (c) an Attention Module, (d) a Downsampling Module, and (e) a Classifier to enhance semantic fidelity.
• Figure 4: A simplified schematic diagram of airborne radar imaging, where $\beta$ and $\varphi$ denote the depression angle and the aspect angle, respectively.
• Figure 5: The embedding results for depression angles of 15 degrees, 30 degrees, and 45 degrees.