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

TL;DR

The paper tackles the gap between conventional super-resolution metrics and downstream SAR ship classification performance. It proposes a classification-aware SR framework comprising three stages—$SR\text{-}I$, $SR\text{-}PT$, and $SR\text{-}FT$—and introduces two task-aligned losses, $L_{combo}$ and $L_{hybrid}$, along with a merged objective $L_{merged} = L_{SR} + L_{CLS}$ to jointly optimize visual fidelity and discriminative power. Using the OpenSARShip dataset with six ship classes and three SR backbones (EDSR, CARN, RCAN) plus five classifiers, the study shows that while $SR\text{-}PT$ improves perceptual quality (PSNR/SSIM), $SR\text{-}FT$ yields the largest gains in classification metrics, with RCAN and VGG16 often providing the best results. The findings validate that task-aligned SR can enhance downstream SAR analytics and motivate extending the framework to detection and segmentation tasks; code and experiments are made available for reproducibility.

Abstract

High-resolution imagery plays a critical role in improving the performance of visual recognition tasks such as classification, detection, and segmentation. In many domains, including remote sensing and surveillance, low-resolution images can limit the accuracy of automated analysis. To address this, super-resolution (SR) techniques have been widely adopted to attempt to reconstruct high-resolution images from low-resolution inputs. Related traditional approaches focus solely on enhancing image quality based on pixel-level metrics, leaving the relationship between super-resolved image fidelity and downstream classification performance largely underexplored. This raises a key question: can integrating classification objectives directly into the super-resolution process further improve classification accuracy? In this paper, we try to respond to this question by investigating the relationship between super-resolution and classification through the deployment of a specialised algorithmic strategy. We propose a novel methodology that increases the resolution of synthetic aperture radar imagery by optimising loss functions that account for both image quality and classification performance. Our approach improves image quality, as measured by scientifically ascertained image quality indicators, while also enhancing classification accuracy.

Figures (8)

• Figure 1: Comparison of attention strategies in super-resolution. Traditional methods apply global attention, while the proposed classification-aware approach uses class label supervision to guide attention toward task-relevant regions, enhancing super-resolution for downstream classification.
• Figure 2: Overview of the proposed classification-aware super-resolution pipeline illustrating the (a) SR-I: inference using ImageNet pretrained SR models, (b) SR-PT: pretraining the SR models on SAR data using Image Quality (IQ) loss functions, and (c) SR-FT: finetuning on SAR data using IQ and classification focused (CLS) loss functions. Evaluation includes PSNR, SSIM, and F1-score. (d) SR Models: SR models used in the pipeline. (e) Feature Extractors: feature extractors used with frozen weights. (f) Classifier Block: classifier layers used for ship classification, where 6 is the number of OpenSARShip classes.
• Figure 3: Overview of the SR-I stage
• Figure 4: Overview of the SR-PT stage
• Figure 5: Architecture of the fine-tuning (SR-FT) stage. The SR model, pre-trained for perceptual quality during SR-PT stage, is fine-tuned using a joint loss function that combines super-resolution loss and task-specific classification loss.