A Classification-Aware Super-Resolution Framework for Ship Targets in SAR Imagery

Technical Summary: A Classification-Aware Super-Resolution Framework for Ship Targets in SAR Imagery

Problem Statement
Synthetic Aperture Radar (SAR) imagery is critical for maritime surveillance and ship classification, yet it faces significant challenges including data scarcity, inherent low resolution, speckle noise, and sidelobe interference. While Super-Resolution (SR) techniques are widely used to reconstruct high-resolution (HR) images from low-resolution (LR) inputs, traditional approaches optimize solely for pixel-level image quality metrics, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). The authors identify a critical disconnect: improvements in these perceptual metrics do not necessarily translate to improved performance in downstream tasks like ship classification. In SAR, where targets are small and dominated by background clutter, standard SR models trained on pixel-wise losses (e.g., L1, L2) often produce over-smoothed results that inadvertently degrade the fine-grained scattering features essential for accurate class discrimination. Prior to this work, no study had introduced a classification-aware SR strategy specifically for SAR ship classification.

Methodology
The paper proposes a unified framework that integrates classification feedback directly into the SR training process. The methodology is structured into three progressive stages, utilizing the OpenSARShip dataset (specifically six classes: Cargo, Tanker, Fishing, Dredging, Passenger, and Tug) and three SR architectures (EDSR, CARN, RCAN) paired with five classification networks (ResNet18/50, VGG16, MobileNetV2, DenseNet121).

1. SR-I (Inference): Baseline evaluation using ImageNet-pretrained SR models to generate HR images from LR inputs, followed by classification to establish a baseline F1-score.
2. SR-PT (Pretraining): SR models are trained on SAR data using Image Quality (IQ) loss functions. Three specific loss functions are evaluated:
   • L1-Loss: Standard pixel-wise reconstruction error.
   • Combo-Loss: A weighted sum ($\alpha=0.5, \beta=0.5$) of PSNR-based and SSIM-based loss terms to guide training toward perceptually and quantitatively improved images.
   • Hybrid-Loss: A weighted combination ($0.7L1 + 0.2LSSIM + 0.1LPSNR$) balancing pixel accuracy, structural integrity, and perceptual quality.
3. SR-FT (Fine-tuning): The pre-trained SR models are fine-tuned using a joint loss function ($L_{merged} = L_{SR} + L_{CLS}$). This stage introduces a classification-guided objective where the SR model is optimized not just for reconstruction fidelity, but also to produce images that yield accurate classification predictions. The classification loss ($L_{CLS}$) is computed as the Mean Squared Error (MSE) between the predicted labels of the SR-generated images and the ground-truth labels of the HR images. Crucially, the feature extraction layers of the classifiers remain trainable during this phase to adapt to the specific SAR domain.

Key Contributions

• Novel Loss Functions: The introduction of "Combo-Loss" and "Hybrid-Loss" designed to guide SR training by integrating multiple image quality metrics.
• Unified Classification-Aware Framework: A methodology that explicitly incorporates classification feedback into the SR training loop, shifting the optimization focus from pure pixel reconstruction to task-relevant feature preservation.
• Empirical Validation: A systematic demonstration that aligning SR optimization with classification metrics yields higher downstream classification accuracy than traditional pixel-focused training, even if it results in slightly lower PSNR/SSIM scores.

Results
Experiments conducted on the OpenSARShip dataset reveal several key findings:

• Divergence of Metrics: There is a clear divergence between image quality metrics and classification performance. Models optimized for PSNR/SSIM (SR-PT) did not always yield the best classification results.
• Superiority of Fine-tuning: The SR-FT stage consistently improved F1-scores across all loss functions compared to the SR-PT stage. For instance, the RCAN model with Combo-Loss improved its F1-score from 61.18% (SR-PT) to 63.41% (SR-FT).
• Loss Function Efficacy: While L1-loss performed adequately, the Combo and Hybrid losses provided a stronger foundation for classification-aware fine-tuning, as they better preserved structural and perceptual cues relevant to ship discrimination.
• Model Performance: Among SR models, RCAN and CARN outperformed EDSR in classification tasks, likely due to their adaptive mechanisms (e.g., channel attention) which help focus on class-relevant features. Among classifiers, VGG16 achieved the highest average F1-score.
• Attention Shift: Grad-CAM visualizations indicate that while SR-I and SR-PT models often focus on background or noisy regions, the SR-FT models produce compact, target-centered activations aligned with the ship structure.

Significance
The paper claims that its primary significance lies in addressing the gap between image reconstruction fidelity and task-specific performance in SAR imagery. It demonstrates that for SAR ship classification, preserving discriminative scattering characteristics is more critical than maximizing pixel-level accuracy. By integrating classification objectives directly into the SR process, the proposed framework produces super-resolved imagery that maintains both visual fidelity and, more importantly, enhances downstream classification accuracy. This approach offers a new paradigm for SAR data processing where the ultimate goal is the improvement of analytical tasks rather than just perceptual quality.