Abstract
In order to improve the stability and rapidity of synthetic aperture radar (SAR) images matching, an effective method was presented. Firstly, the adaptive smoothing filtering was employed for image denoising in image processing based on Wallis filtering to avoid the follow-up noise is amplified. Secondly, feature points were extracted by a simplified SIFT algorithm. Finally, the exact matching of the images was achieved with these points. Compared with the existing methods, it not only maintains the richness of features, but also reduces the noise of the image. The simulation results show that the proposed algorithm can achieve better matching effect.
Introduction
Image registration—the process of overlaying two or more images of the same scene taken at different times or from different sensors—is a fundamental step in remote sensing. While the Scale-Invariant Feature Transform (SIFT) is considered a gold standard for optical imagery, it often fails when applied to Synthetic Aperture Radar (SAR) data.
The primary challenge is speckle noise, a strong multiplicative noise inherent to SAR imaging. Speckle creates "false" gradients that confuse SIFT’s feature detectors, leading to a low number of correct matches and a high rate of outliers. Furthermore, SAR images often suffer from low local contrast, making it difficult to extract the distinctive keypoints SIFT relies on. This paper proposes a robust registration pipeline that combines adaptive smoothing and Wallis filtering with a simplified SIFT algorithm to overcome these limitations.
Methodology
The proposed method utilizes a three-stage preprocessing and matching pipeline to ensure reliable feature extraction from noisy SAR data.
1. Adaptive Smoothing Filtering
To tackle speckle without blurring important structural details, the authors use an iterative adaptive smoothing filter. This technique uses local gradients to adjust filter weights: it smooths heavily in uniform areas to suppress noise but preserves edges where the gradient is high. This ensures the structural integrity of the image is maintained for subsequent steps.
2. Wallis Filtering
Once the noise is suppressed, the image undergoes Wallis filtering. This is a local contrast enhancement technique that maps the local mean and variance of the image to specific target values. This "stretches" the texture information, making fuzzy features more distinct for feature extraction. Crucially, by placing the adaptive smoothing filter before the Wallis filter, the authors ensure they are only enhancing real structural information rather than amplifying noise.
3. Simplified SIFT & RANSAC
With the preprocessed image, a simplified SIFT algorithm is used to detect and describe feature points. Finally, the RANSAC (Random Sample Consensus) algorithm is employed to filter out mismatches and calculate the final transformation between the two images.
Experimental Results
The study evaluated the method using three pairs of 200 x 200 pixel SAR images representing typical challenges like serious speckle and similar texture structures.
Quantitative Performance
The proposed method consistently outperformed traditional SIFT across all test cases in both the quantity and quality of matches:
| Metric | Data 1 (Trad. / Prop.) | Data 2 (Trad. / Prop.) | Data 3 (Trad. / Prop.) |
|---|---|---|---|
| Matching Points | 50 / 110 | 228 / 329 | 83 / 305 |
| Correct Matches | 45 / 108 | 203 / 319 | 71 / 278 |
| Correct Rate (%) | 88% / 98% | 89% / 97% | 86% / 91% |
Beyond matching statistics, the adaptive smoothing improved image quality metrics significantly. The Peak Signal-to-Noise Ratio (PSNR) reached up to 38.78 dB, while Structural Similarity (SSIM) values remained high (0.94–0.96), indicating that the denoising process did not destroy the underlying structure of the original image.
Conclusion and Future Work
The integration of adaptive smoothing and Wallis filtering provides a robust foundation for SAR image registration. By suppressing noise before enhancing contrast, the method creates the sharp local extrema in the Gaussian scale-space that SIFT needs to function effectively. While the iterative nature of adaptive smoothing adds computational overhead, the gains in registration accuracy are substantial. Future work could explore automated parameter tuning to make the method more robust across varying noise statistics and larger-scale datasets.
Summary updated