Abstract
Synthetic Aperture Radar (SAR) is a critical technology for all-weather remote sensing, yet it is frequently hampered by speckle noise and complex background clutter. Traditional detection algorithms, such as Constant False Alarm Rate (CFAR), often struggle in these environments, leading to high false-positive rates. This paper introduces a robust three-stage target detection framework that mimics human visual attention. The method integrates shadow-based object proposals, spatial autocorrelation saliency analysis using $G$ statistics, and one-class classification. Experimental results on the MSTAR database demonstrate that the proposed approach achieves a precision of 94.74% and significantly reduces computational overhead compared to existing saliency models.
Introduction
Synthetic Aperture Radar (SAR) provides high-resolution imagery regardless of lighting or weather conditions. However, the coherent nature of SAR leads to "speckle noise," a granular interference that complicates automated target detection. Conventional methods typically prioritize high-intensity pixels, which often results in false alarms in cluttered scenes where non-target objects may also appear bright.
A key insight of this research is the utilization of shadows. In radar imaging, physical objects cast distinct shadows that are often more stable and reliable cues for localization than the targets themselves. By focusing on the relationship between targets and their shadows, the search space for detection can be drastically reduced.
Methodology
1. Shadow Proposal Stage
The framework begins by identifying potential target locations through their shadows. Using dual-threshold segmentation and morphological filtering, the algorithm extracts shadow regions. A Mean Shift algorithm is then employed to model the spatial offset between a shadow and the object casting it, allowing the system to generate "proposal chips"—small image windows likely to contain a target.
2. GSST Saliency Detection
Within the proposal chips, the system identifies salient regions using the $G$ statistic, a measure of local spatial autocorrelation. Unlike traditional methods that define saliency by brightness alone, this approach defines it as the density of high-intensity pixels.
By calculating $G$ statistics, the algorithm can distinguish between compact, "clustered" targets and random noise or dispersed clutter. A subsequent significance test ensures that only statistically relevant salient regions are passed to the final classification stage.
3. OC-SVM Screening
The final stage involves distinguishing true targets (e.g., vehicles) from salient clutter. The researchers utilize a One-Class Support Vector Machine (OC-SVM). This choice is strategic for SAR applications where positive samples (targets) are available, but modeling the infinite variety of potential "background" anomalies is impractical. The OC-SVM learns the feature distribution of known targets and rejects outliers as clutter.
Experimental Results
The framework was evaluated using the MSTAR (Moving and Stationary Target Acquisition and Recognition) benchmark dataset, featuring X-band radar images of military vehicles.
- Saliency Performance: The GSST saliency method achieved an $F_\beta$-measure of 90.71%, outperforming the Graph-Based Visual Saliency (GBVS) and Region Contrast (RC) models.
- Detection Accuracy: In complex test scenes, the method reached 94.74% precision and 90% recall.
- Efficiency: The system is highly optimized for real-time applications, with the saliency detection stage requiring only 0.087 seconds per image.
- Robustness: High performance was maintained in environments with a Signal-Clutter-Ratio (SCR) as low as 4dB.
Conclusion
This research presents a computationally efficient and precise pipeline for SAR target detection. By leveraging the "universal property" of shadows and employing spatial autocorrelation for saliency, the method effectively mitigates the impact of speckle noise and clutter. While deep learning models like Faster R-CNN may offer higher robustness in extremely low SCR conditions, the proposed framework provides a superior balance of speed and accuracy for standard remote sensing tasks. Future work will explore the integration of deep features into the one-class classification stage to further enhance performance in ultra-noisy environments.
Summary updated