PaperlayerView exact revision on arXiv

Source-linked AI summary

ProSR: Semantic-Prototype-Guided Discrete Modeling for Physically Consistent SAR Super-Resolution

Byoungwoo Kim, Munchurl Kim

arXiv:2609.02377v1cs.CV

TL;DR

SAR super-resolution needs to recover high-resolution imagery without losing physically meaningful scattering, while high-resolution acquisition and labeled SAR data remain limited. ProSR uses semantically guided discrete token prediction, self-supervised priors, detail encoding, and prototype-map-guided attention; on a 0.25 m Umbra benchmark, it reports superior realism and structural accuracy. Its stated scope includes challenges with sub-resolution targets, SSL capacity, and validation-set diversity.

  • Problem

    High-resolution SAR imagery is costly to acquire, and smooth generative ISR models can distort coherent scattering statistics and produce structures inconsistent with SAR physics.

  • Method

    ProSR performs semantically guided discrete token prediction in a quantized detail latent space, using SSL-derived priors, SADE, SPMG, and PMGA to guide reconstruction.

  • Results

    ProSR achieves superior physical realism and structural accuracy while suppressing structural hallucinations on the 0.25 m Umbra SAR ISR benchmark.

  • Takeaways & Limitations

    The framework preserves sharp, physically consistent scattering without requiring manually labeled data for the reconstruction guidance.

  • Takeaways & Limitations

    The reported limitations include sub-resolution targets, SSL capacity constraints, and validation-set diversity; the prototype design also found K > 3 redundant or inactive.

Abstract

from arXiv · show

High-resolution Synthetic Aperture Radar (SAR) imagery is critical for precision analysis such as automatic target recognition, yet its acquisition is costly. Although generative image super-resolution (ISR) models offer a promising alternative, current smooth-approximation based diffusion frameworks often struggle to preserve the coherent scattering statistics, causing stochastic structural distortions that are less consistent with real SAR physics. To address this, we propose Semantic Prototype-Guided Super-Resolution (ProSR), reformulating SAR ISR as a semantically-guided discrete token prediction task within a quantized latent space. By mapping signal features to discrete scattering primitives, ProSR preserves the impulsive nature of SAR without over-smoothing. Furthermore, we integrate a Self-Supervised Learning backbone into SAR ISR to extract label-free semantic priors, overcoming label scarcity. Guided by these priors, we introduce Semantic-Aligned Detail Encoding to decouple high-frequency signals into discrete scattering primitives. In parallel, the Semantic Prototype Map Generator explicitly constructs semantic prototype maps, allowing Prototype-Map-Guided Attention to route the information flows within identical categories and mitigate inter-class interference. To validate our approach, we present a large-scale 0.25m resolution benchmark from the Umbra Open Dataset. Experimental results show ProSR achieves superior visual quality while preserving essential scattering characteristics required for practical SAR applications.

1 Introduction

ProSR addresses SAR super-resolution’s difficulty preserving physically meaningful scattering under smooth generative modeling by combining discrete token prediction with label-free semantic guidance. It introduces semantic detail encoding, prototype-map-guided attention, and a 0.25 m Umbra benchmark, reporting improved realism and scattering preservation.

  • High-resolution SAR imagery supports precision applications such as automatic target recognition and infrastructure monitoring, but acquiring it is costly.
  • Smooth-approximation diffusion models can produce stochastic structural distortions in SAR by interpolating between disjoint data modes.
  • ProSR reformulates SAR ISR as semantically guided discrete token prediction over physically valid codebook entries rather than smooth density approximation.
  • Self-supervised semantic guidance addresses scarce labels through Semantic-Aligned Detail Encoding and a semantic-aware perceptual loss.
  • SPMG constructs spatial-semantic maps that guide PMGA toward semantically consistent regions, reducing inter-class signal leakage during reconstruction.
  • 0.25 m resolution benchmark from the Umbra Open Dataset enables evaluation, while ProSR reports superior visual realism and preservation of essential scattering characteristics.

2 Related Works

Prior SAR super-resolution methods progressed from optimization and generative models toward sharper reconstructions, but diffusion can still cause physically harmful mode interpolation. ProSR extends discrete modeling and self-supervised representation learning into SAR reconstruction.

  • Early pixel-wise SISR architectures often produced over-smoothed results, whereas GANs improved sharpness but frequently introduced unrealistic hallucinations.
  • Diffusion models improved perceptual fidelity and training stability but rely on smooth data-distribution approximations that can generate structural hallucinations.
  • Discrete representation learning replaces continuous latent values with categorical selections from a finite codebook, discouraging blurred intermediate values.
  • SAR ISR must model speckle noise and complex electromagnetic interactions because realistic scattering patterns support operational analyses such as ATR.
  • Self-supervised learning offers label-free representation learning, but its SAR use has largely focused on classification and semantic segmentation.
  • ProSR integrates SSL-derived semantic priors into low-level reconstruction, addressing an underused role for SSL features in SAR ISR.

3 Methodology

ProSR uses a three-stage pipeline that separates semantic structure from high-frequency SAR details, discretizes scattering representations, and guides masked token generation with semantic prototypes. The design combines self-supervised semantic priors, spatial-semantic maps, and restricted attention to produce physically aligned SAR details.

  • Overall architecture: ProSR comprises Semantic-Aligned Detail Encoding, Semantic Prototype Map Generation, and Prototype-Guided Masked Generative Modeling.Stage 1 separates detail and structural representations; Stage 2 constructs spatial-semantic guidance; Stage 3 generates high-resolution tokens.
  • Stage 1: SADE: SADE computes a residual detail feature z_det = z − F_sem and quantizes it into codebook entries, decoupling high-frequency details from low-frequency semantic context.The detail encoder maps HR imagery to z, while the semantic pathway derives F_sem from LR input using a self-supervised backbone and adapter.
  • Stage 1: SADE: A stochastic two-branch decoder reconstructs LR structure from F_sem alone with probability p and HR imagery from ẑ_det + F_sem with probability 1 − p.This strategy encourages F_sem to preserve semantic structure independently while reserving ẑ_det for high-frequency scattering signatures.
  • Stage 1: SADE: The detail codebook uses Gumbel vector quantization and KL regularization to support differentiable training, prevent codebook collapse, and encourage diverse scattering primitives.The KL objective maximizes codebook entropy, while the Stage 1 perceptual loss uses a pretrained SAR self-supervised model to capture domain-specific signatures without labels.
  • Stage 3: PMG: PMG predicts masked HR tokens iteratively as categorical codebook selections conditioned on masked details, semantic features, and the spatial-semantic map.This discrete formulation keeps generated signals within physically valid scattering primitives and suppresses mode interpolation.
  • Stage 3: PMG: PMGA restricts cross-attention to semantically consistent regions, suppressing inter-class feature leakage while preserving sharp intensity distributions and SAR speckle statistics.Unassigned regions bypass cross-attention to block background interference, and the resulting iterative refinement produces details aligned with the physical scattering layout.

4 Experimental Results

Experiments evaluate ProSR on a standardized Umbra-based SAR benchmark using amplitude-domain fidelity, scattering, statistical, and perceptual metrics. Results favor detail-focused quantization and semantic prototype guidance for structural and statistical realism, with downstream ATR gains.

  • Benchmark and Evaluation: The benchmark uses 0.25 m HR references and 1.0 m LR counterparts derived from Umbra X-band SLC data.Training used 256×256 crops and sliding-window inference with stride 128.
  • Benchmark and Evaluation: Evaluation combines PSNR and SSIM with IW-SSIM, HaarPSI, |∆TCR|, FID, Density, Coverage, LPIPS, and DISTS.The complementary metrics target scattering fidelity, structural coherence, statistical realism, and perceptual similarity.
  • Quantitative and Qualitative Results: ProSR dominates Target and Structural Integrity metrics while also excelling in FID, Density, and Coverage, despite diffusion baselines achieving higher PSNR/SSIM.The reported comparison distinguishes pixel-aligned fidelity from SAR-specific scattering and statistical realism.
  • Downstream Utility: ProSR achieves 88.39% ATR accuracy, compared with 23.87% for oversampled LR and 84.18% for the strongest baseline.The classifier was fixed and trained solely on original HR images, yielding a 4.21%p gain over the strongest baseline.
  • Ablation Studies: Detail-VQ outperforms standard HR-VQ and avoids the structural errors introduced by Dual-VQ’s quantization of LR components.The detail-only codebook decouples high-frequency representation from LR distortions, reducing over-sharpening while preserving structural and statistical fidelity.
  • Ablation Studies: A codebook size of N = 1024 is selected because N = 1024 and 2048 have comparable LPIPS and SSIM, while N = 1024 achieves highest PSNR and simplifies Stage 3 classification.N = 512 is insufficient for complex textures.
  • Ablation Studies: PMGA reduces FID by 1.69 and improves Density to 0.9741 and Coverage to 0.8592 while discriminating scattering centers.The reported heatmaps show pixel-wise averaged attention weights associated with target, clutter, and shadow categories.

5 Conclusion

The conclusion presents ProSR as a semantically guided discrete-token framework for SAR ISR that addresses smooth-model distortions and labeled-data scarcity. Its SADE-defined detail latent space and SSL-derived semantic priors support physically consistent scattering reconstruction on a 0.25 m benchmark.

  • Conclusion: ProSR reformulates SAR ISR as semantically guided discrete token prediction within a discrete detail latent space defined by SADE.The framework targets stochastic structural distortions associated with smooth-approximation-based diffusion models.
  • Conclusion: The SSL backbone extracts label-free semantic priors, while SPMG constructs Msem and PMGA routes information within consistent categories.The conclusion connects these components to reduced inter-class confusion and sharp, physically consistent scattering restoration without manually labeled data.
  • Conclusion: Experiments on the 0.25 m resolution benchmark validate superior physical realism and structural fidelity for ProSR.The supplied conclusion passage states this validation without reporting a specific metric.

- Supplementary Material -

The supplementary material identifies the paper’s authors and their institutional affiliation. It also provides their contact email addresses.

  • Authorship: The paper lists Byoungwoo Kim and Munchurl Kim as authors.
  • Affiliation: The authors are affiliated with the Korea Advanced Institute of Science and Technology in Daejeon, Republic of Korea.
  • Contact: The listed contact addresses are {quddn826, mkimee}@kaist.ac.kr.

Supplementary Overview

The supplementary material organizes additional dataset analysis, experiments, theoretical discussion, efficiency analysis, and failure-case documentation. Its limitations section addresses sub-resolution targets, SSL capacity, and validation-set diversity.

  • Dataset Analysis: Section A analyzes SAR dataset characteristics and data distributions.
  • Additional Experiments: Section B adds qualitative results, MSTAR ATR evaluation, DiT and prototype ablations, and AE stochastic-training analysis.
  • Additional Discussions: Section C discusses diffusion-model toy theory and factors affecting SSIM in generative SAR super-resolution.
  • Model Complexity and Efficiency: Section D analyzes parameter counts, FLOPs, and inference runtime.
  • Failure Cases and Limitations: Section E documents challenges with dense sub-resolution targets, SSL capacity constraints, and validation-set diversity.

A Dataset Analysis

The benchmark contains geographically diverse SAR scenes spanning five semantic categories and supports evaluations of reconstruction, scattering fidelity, and downstream ATR transfer. Analyses compare ProSR with continuous and diffusion-based baselines, including prototype-token sensitivity and qualitative performance across environments.

  • Dataset composition: The curated dataset contains 502 SAR images categorized into Airport, Urban, Industrial, Port, and Natural scenes.Categories were assigned using predominant environmental context and site metadata.
  • Dataset composition: The dataset covers diverse geographic locations, environmental conditions, categories, and incidence angles from 10°–50°.Global distribution and category-by-incidence-angle statistics are reported in the dataset analysis.
  • Dataset construction: Uniform regions were selectively discarded to balance information-dense target structures against homogeneous speckle areas and reduce generative mode collapse.This preprocessing emphasizes structurally informative regions during training.
  • Evaluation protocol: The MSTAR ATR transfer experiment evaluates whether reconstructed images preserve HR scattering characteristics under an 80%/20% train/test split.ISR models pretrained on the benchmark were fine-tuned for 100 epochs before evaluation.
  • Modeling analysis: The controlled continuous-versus-discrete comparison reports lower FID and LPIPS for ProSR than the DiT-style variant: FID 23.70 vs. 40.43 and LPIPS 0.3010 vs. 0.3211.The analysis attributes this gap to discrete modeling better preserving distinct scattering structures.
  • Prototype-token sensitivity: Three prototype tokens separate target, clutter, and shadow more clearly than fewer or redundant tokens.K = 2 under-segments, while K ≥4 introduces redundant tokens that reduce semantic confidence.

B.5 Sensitivity Analysis of Relative Masking Threshold (γ)

The masking-threshold analysis evaluates how relative threshold γ affects semantic prototype maps and class-specific information exchange. γ = 0.7 provides the reported balance between semantic purity and structural connectivity, whereas extreme settings create over-expansion or omission.

  • Evaluation setup: The threshold study visualizes semantic prototype maps alongside LR images, with clutter masks using the top c = 30% of pixels ranked by global clutter score.The selected clutter coverage is intended to isolate stable scattering signatures.
  • Baseline masking: Naive argmax misclassifies clutter within target regions because it lacks a filtering mechanism for localized scattering peaks.The resulting masks fail to represent distinctive clutter characteristics.
  • Mask quality: Semantic contamination mixes heterogeneous target and shadow features within a single mask, hindering class-specific information exchange.This weakens separation of class-specific characteristics.
  • Mask quality: Insufficient valid structural segments reduce intra-class information exchange and produce fragmented, disconnected scattering representations.The affected masks fail to capture complete scene geometry.
  • Threshold selection: γ = 0.7 provides the proper balance for informative class-specific interactions, preserving semantic purity and structural connectivity.This enables high-fidelity reconstruction using distinct scattering characteristics.

C.1 Toy analysis of close-mode separation in diffusion models

The toy analysis explains why closely spaced SAR scattering modes can remain merged during diffusion denoising until low noise levels. It motivates ProSR’s categorical codebook selection and also documents practical limits from misalignment, latency, and severe sub-resolution degradation.

  • Close-mode separation: For the toy two-mode distribution, the score changes from a merged peak at σ_t > a to a valley separating modes at σ_t < a.Thus, close peaks become distinguishable only below the noise scale associated with their separation.
  • Close-mode separation: Finite sampling steps and score-approximation errors may under-resolve late-emerging separation, leaving probability mass between valid peaks and smearing scattering responses.The issue is most relevant when distinct SAR scatterers are closely spaced.
  • Discrete alternative: ProSR uses hard categorical codebook selection to reduce interpolation between distinct structural codes and preserve learned scattering patterns.Its discrete formulation is intended to maintain sharper impulse-like signatures.
  • Metric interpretation: SSIM can penalize realistic generative reconstructions when sharply localized responses shift by only 1–2 pixels or when speckle realizations differ.Pixel alignment and speckle stochasticity make this metric counterintuitive for SAR ISR.
  • Efficiency trade-off: ProSR has an 85.83M generative core and fewer FLOPs than ResShift but higher inference runtime because global attention supports long-range dependencies during decoding.The paper treats this latency as a trade-off for reconstruction accuracy.
  • Failure cases: Dense sub-resolution targets can become indistinguishable from speckle after coherent phase mixing, causing ProSR and other baselines to produce fragmented outputs.In these cases, geometric information is irreversibly degraded before reconstruction.

E.2 Limitations

The paper identifies limitations in semantic-feature capacity and validation-set diversity. These constraints motivate richer SSL representations and broader benchmark sources.

  • Semantic guidance: ProSR’s reconstruction precision depends on the representational richness of the SSL latent space used for semantic maps and cross-attention values.Local semantic-map inaccuracies may cause minor misalignments.
  • Dataset diversity: The validation set is concentrated on specific categories, predominantly port patches, despite using complex non-overlapping environments.The authors plan broader sources for more balanced benchmarks.
Loading 2609.02377v1…