Destriping of Remote Sensing Images by an Optimized Variational Model

Fei Yan, Siyuan Wu, Qiong Zhang, Yunqing Liu, Haonan SunSensors 2023 · DOI

TL;DR

This paper introduces a novel variational model for removing stripe noise from remote sensing images. By incorporating an optimized lp quasinorm and a fast ADMM solver, the method effectively removes stripes while preserving fine image details, outperforming existing approaches in both simulated and real-world scenarios.

What Problem Does This Paper Address?

Remote sensing images, crucial for applications like environmental monitoring and disaster assessment, are often plagued by stripe noise. This noise, arising from sensor imperfections and mechanical issues, significantly degrades image quality and hinders subsequent analysis. While various destriping methods exist, they often struggle with complex non-periodic stripes, lead to loss of fine details, or introduce artifacts. This paper aims to address these limitations by developing a more robust and detail-preserving destriping technique.

Methods

The core of this work is a new unidirectional variational model designed to tackle horizontal stripe noise. Key innovations include:

  • Optimized Variational Model: The model incorporates directional characteristics and structural sparsity of stripe noise. It uses an lp quasinorm (with 0 < p < 1) to better characterize the global sparsity of stripe noise and the local sparsity of vertical image gradients. This choice of norm is shown to induce sparser solutions than traditional l1 or l2 norms. Additionally, an l0 norm is used to capture local sparsity in the horizontal gradient, further preventing detail loss.
  • Fast Alternating Direction Method of Multipliers (Fast ADMM): To efficiently solve the resulting non-convex optimization problem, the authors employ a fast ADMM algorithm. This approach significantly reduces computation time compared to traditional ADMM, making it suitable for large-scale datasets. The algorithm involves introducing intermediate variables and solving subproblems for each variable iteratively.

Datasets

The authors evaluated their method on a variety of datasets:

  • Simulated Data: Six remote sensing images were used, including MODIS data (bands 31 and 22), Tiangong-1 hyperspectral data, Washington DC Mall hyperspectral data, and VIIRS hyperspectral data. Stripe noise of varying intensities and ratios (periodic and non-periodic) was synthetically added to these images.
  • Real Data: Four MODIS dataset images were used, affected by different types and levels of stripe noise (periodic and non-periodic).

Key Insights and Ideas

  • The Power of lp Quasinorm: The lp quasinorm (0 < p < 1) offers greater flexibility than l1 or l2 norms in characterizing sparsity. Its contours more closely approximate the coordinate axes, leading to sparser solutions and better preservation of image details, especially in the presence of complex stripe noise.
  • Directional Sparsity is Crucial: Recognizing that stripe noise primarily affects gradients in specific directions (vertical for horizontal stripes), the model leverages directional sparsity constraints. This targeted approach is more effective than general sparsity assumptions.
  • Fast ADMM for Efficiency: The adoption of Fast ADMM is a significant practical improvement, drastically reducing computation time and making the method more viable for large-scale remote sensing applications.
  • Robustness to Noise Levels: The proposed method demonstrates robustness across various stripe noise intensities and types, outperforming existing methods in many scenarios.

Key Results

  • Qualitative Improvements: Visual inspection of destriped images shows that the proposed method effectively removes stripes with fewer artifacts and less blurring of fine details compared to WAFT, WLS, UTV, SAUTV, GSLV, and RBSUTV. This is particularly evident in magnified regions of interest.
  • Quantitative Performance:
    • Simulated Data: The method achieved superior PSNR and SSIM values compared to most other methods, especially under higher stripe noise levels. For example, in Table 1, the proposed method consistently shows competitive or superior PSNR and SSIM scores.
  • Real Data: Objective evaluation metrics (MRD and ICV) in Table 3 indicate that the proposed method outperforms others in most cases, demonstrating its effectiveness and robustness on real-world data.
  • Computational Efficiency: Table 4 shows that Fast ADMM is significantly faster than standard ADMM, with the speedup increasing for larger image sizes. For a 700x700 image, Fast ADMM is over three times faster.

Limitations and Future Work

  • Parameter Selection: The authors acknowledge that selecting optimal regularization parameters (λ1 and λ2) can be challenging and time-consuming, often requiring extensive experimentation. While they suggest a greedy approach, adaptive parameter selection is a future research direction.
  • Small-Scale Random Stripes: The method struggles with very small-scale random stripe noises that disrupt overall sparsity and low-rank properties, sometimes failing to completely remove them without compromising overall image quality (as shown in Figure 23).
  • Future Directions: The authors propose extending the method to other stripe noise removal applications beyond remote sensing. They also plan to explore additional stripe and image component characteristics and integrate deep learning methods with traditional approaches for even more effective destriping.

Summary updated

BibTeX
@article{yan2023destriping,
  author = {Fei Yan and Siyuan Wu and Qiong Zhang and Yunqing Liu and Haonan Sun},
  title = {{Destriping of Remote Sensing Images by an Optimized Variational Model}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23177529},
}
Figure