Abstract
Stripe noise is a persistent degradation in satellite imagery, often becoming oblique after geometric registration and geo-rectification. Traditional destriping methods typically require rotating the image to align stripes with the axes, a process that introduces resampling artifacts and loss of sub-pixel accuracy. This paper introduces a robust optimization-based framework utilizing an "Oriented Variation" (OV) model. By estimating the stripe angle and adapting differential operators accordingly, the proposed method removes diagonal stripes directly without damaging the underlying image structure. Experimental results on simulated and real-world datasets demonstrate superior performance in terms of both quantitative metrics and visual fidelity.
Introduction
Stripe noise in remote sensing images is primarily caused by inconsistent responses among the multiple detectors used in satellite imaging systems. While vertical or horizontal striping is common in raw data, "higher-level" satellite products often undergo geometric registration. This process rotates the original stripes, resulting in oblique stripes that are significantly harder to remove using standard algorithms.
Existing solutions often involve a three-step process: rotating the image to make stripes axis-aligned, applying a standard destriping filter, and rotating the image back. However, this double-rotation involves resampling that can introduce "jagged" patterns and blur fine geographical details. This research proposes a method to remove these diagonal stripes directly in their original orientation.
Methodology
The proposed Oriented Variation (OV) model avoids image rotation by instead rotating the mathematical operators used to detect and remove noise. The framework consists of three primary stages:
1. Stripe Orientation Estimation
To accurately identify the angle of the stripes, the authors employ a "denoising the noise" strategy. First, a guided filter is used to suppress the complex terrain information, which often obscures the noise pattern in the frequency domain. A Fast Fourier Transform (FFT) is then applied to the filtered image. Because stripes are periodic, they appear as distinct peaks in the Fourier spectrum. By identifying these peaks, the model can calculate the stripe angle ($\theta$) with high precision.
2. Oriented Variation (OV) Constraint
Standard Total Variation (TV) models use horizontal and vertical derivatives. The OV model expands this by using a set of candidate directional derivatives. Once the stripe angle is estimated, the model selects the derivative operator that most closely aligns with the stripe direction. This enforces "along-stripe smoothness," ensuring that pixels within a single stripe are treated as a continuous unit while preserving the sharp edges of the underlying terrain.
3. Joint Optimization
The final energy function combines three constraints to balance image restoration and noise removal:
- Total Variation (TV): Maintains piece-wise smoothness in the clean image.
- Oriented Variation (OV): Captures the specific directional characteristics of the stripes.
- $L_1$-norm: Models the global sparsity and distribution of the stripe noise.
This complex optimization problem is solved efficiently using the Alternating Direction Method of Multipliers (ADMM).
Experimental Results
The model was rigorously tested on simulated data (from Terra/Aqua MODIS, HYDICE, and Hyperion) and real-world noisy imagery (from HJ1A/HIS, Sentinel 2A, and LANDSAT 5 TM).
- Accuracy: The orientation estimation stage achieved an absolute error of less than 0.7° across all test cases.
- Quantitative Performance: The OV model consistently outperformed existing methods like Wavelet-based Oblique Stripe Removal (WAOR) in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).
- Visual Quality: Unlike rotation-based methods, the OV framework successfully retained delicate features such as thin river networks and sharp mountain ridges without introducing jagged artifacts.
- Efficiency: The algorithm is highly scalable, processing a 2000x2000 pixel image in approximately two minutes on standard hardware.
Conclusion
The Oriented Variation model provides a flexible and robust solution for the persistent problem of oblique stripe noise. By rotating the differential operators rather than the image pixels, it preserves sub-pixel accuracy and fine textures. While the current model assumes a global stripe orientation, future iterations aim to implement local-based estimation to handle complex, non-parallel striping patterns found in wide-swath satellite imagery.
Summary updated