WE-KAN: SAR Image Rotated Object Detection Method Based on Wavelet Domain Feature Enhancement and KAN Prediction...

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Synthetic Aperture Radar (SAR) imagery is crucial for military reconnaissance and disaster monitoring, demanding high detection accuracy.
• Existing rotated object detection methods struggle with SAR imagery's speckle noise and complex backgrounds, limiting precision.
• Accurate detection of elongated objects like ships and bridges in SAR scenes requires precise orientation angle prediction.

Purpose of the Study:

• To propose a novel rotated object detection framework, WE-KAN, specifically designed for complex SAR imagery.
• To enhance SAR object detection accuracy by integrating wavelet features and a Kolmogorov-Arnold Network (KAN).
• To improve robustness against speckle noise and enhance perception of fine object structures in SAR scenes.

Main Methods:

• Incorporated wavelet domain features from SAR grayscale images into the backbone network.
• Fused wavelet and image features using a proposed attention module.
• Designed a KAN-based angle predictor for sensitive angle regression and employed a joint loss function (RIoU + Gaussian distance) for rotated bounding box regression.

Main Results:

• Achieved an AP50 of 70.1 and a mAP of 35.9 on the large-scale public RSAR dataset.
• Significantly outperformed existing baselines under identical training schedules and backbone networks.
• Demonstrated improved robustness to noise and enhanced perception of fine object structures.

Conclusions:

• The proposed WE-KAN framework is effective and robust for detecting dense, small, and highly oriented objects in complex SAR scenes.
• The integration of wavelet features and KAN significantly boosts rotated object detection performance in SAR imagery.
• WE-KAN offers a powerful solution for high-precision object detection in critical SAR applications.