Compressive SAR Imaging with Joint Sparsity and Local Similarity Exploitation

Fangfang Shen, Guanghui Zhao, Guangming Shi, Weisheng Dong, Chenglong Wang, Yi Niu

TL;DR

This paper introduces a novel compressive SAR imaging approach that adapts to complex scenes by using a weighted autoregressive model to exploit local image structure. This method significantly improves image quality and reduces artifacts compared to existing techniques, especially in challenging scenarios.

What Problem Does This Paper Address?

Synthetic Aperture Radar (SAR) is a powerful remote sensing tool, but achieving high-resolution images requires significant data, posing challenges for data acquisition and processing. Compressive Sensing (CS) offers a solution by reconstructing high-resolution images from fewer measurements, assuming the image can be sparsely represented in a fixed basis (like wavelets or DCT). However, this assumption breaks down in complex scenes where local image structures vary significantly. Fixed sparse representations lack adaptivity, leading to artifacts and degraded image quality. This paper addresses this limitation by proposing a CS-SAR method that dynamically adapts its sparse representation to local image characteristics.

Methods

The core of the proposed approach is an Adaptive Sparse Representation (ASR) scheme built upon a weighted autoregressive (AR) model.

  • Autoregressive (AR) Model: This model captures local image structure by representing a pixel's value as a linear combination of its neighbors, plus a random perturbation. The authors acknowledge that traditional AR models assume local stationarity, which may not hold in complex SAR scenes.
  • Structural Similarity Probability (SSP): To address the non-stationarity, the paper introduces SSP to weigh the influence of neighboring pixels. SSP quantifies the similarity between a pixel's local structure and that of its neighbors. Pixels with similar structures receive higher weights, allowing the AR model to adapt to local variations. This is mathematically formulated using a Gaussian function based on the Euclidean distance between local structure vectors.
  • Weighted AR Model: By incorporating SSP as weights into the AR model, a weighted AR model is derived. This model is more adaptive to the varying local waveforms and structures within a SAR image.
  • Joint Optimization: Since the parameters of the weighted AR model depend on the unknown SAR image itself, the paper proposes a joint optimization scheme. This involves iteratively performing SAR imaging and updating the weighted AR model parameters. This is framed as a convex optimization problem, solved using an Augmented Lagrange Multiplier (ALM) technique and conjugate gradient (CG) algorithm for efficiency.

Datasets

The paper evaluates its approach using both synthetic and real SAR data.

  • Synthetic Data: Used for initial validation and comparison with existing methods under controlled conditions.
  • Real SAR Data: Acquired by an airborne SAR, this data is used to demonstrate the approach's effectiveness in a practical scenario, including handling phase errors.

No new benchmarks are introduced; rather, the performance is compared against established methods on existing datasets.

Key Insights and Ideas

  • Adaptivity is Key for Complex Scenes: The paper strongly argues that fixed sparse representations are insufficient for complex SAR scenes. The ability to adapt the sparse representation to local image structure is crucial for high-quality reconstruction.
  • Structural Similarity as a Powerful Weighting Mechanism: Using SSP to guide the AR model's adaptation is a novel and effective way to capture local image properties. It allows the model to prioritize information from structurally similar regions, leading to better representation.
  • Joint Optimization is Necessary: The interdependence between the image content and the adaptive model parameters necessitates a joint optimization strategy. This iterative approach allows both the image and the model to refine each other.
  • Improved Artifact Suppression: The adaptive nature of the proposed method leads to a significant reduction in artifacts that plague traditional CS-SAR approaches, particularly in areas with complex textures like roads and terraces.

Key Results

  • Superior Image Quality: The proposed ASR-CS-SAR approach consistently outperforms conventional CS-SAR methods (using wavelet, DCT, and TV bases) in terms of both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) on both synthetic and real SAR data.
  • Significant PSNR/SSIM Gains: For low sampling modes, the average gain over competing methods can be up to 4 dB in PSNR and 0.02 in SSIM. These gains are further amplified with increased sampling rates.
  • Effective Artifact Reduction: Visual comparisons show that the ASR-CS-SAR method produces sharper and clearer images with significantly suppressed artifacts, particularly in complex regions.
  • Robustness to Sampling Rates: The approach demonstrates strong performance across various sampling rates, maintaining its superiority as sampling rates increase.
  • Effective Phase Error Compensation: When applied to real SAR data with phase errors, the proposed method, in conjunction with an autofocus strategy, yields a more focused and accurate image compared to other methods.

Limitations and Future Work

  • Computational Complexity: While the paper employs techniques like CG to mitigate complexity, the iterative nature of the joint optimization and the calculation of SSP can still be computationally intensive, especially for very large images. The paper mentions that matrix-vector multiplication in CG can be parallelized, suggesting future work in hardware acceleration.
  • Window Size Selection: The AR model's performance is influenced by the local window size. An empirical 3x3 window was used, but the paper suggests this is an empirical choice to balance efficiency and robustness. Further investigation into optimal window size selection or adaptive windowing could be beneficial.
  • Generalization to Other Imaging Modalities: While effective for SAR, the adaptability of the ASR framework might be explored for other imaging modalities where local structural variations are significant.

Summary updated

Abstract

Compressive sensing-based synthetic aperture radar (SAR) imaging has shown its superior capability in high-resolution image formation. However, most of those works focus on the scenes that can be sparsely represented in fixed spaces. When dealing with complicated scenes, these fixed spaces lack adaptivity in characterizing varied image contents. To solve this problem, a new compressive sensing-based radar imaging approach with adaptive sparse representation is proposed. Specifically, an autoregressive model is introduced to adaptively exploit the structural sparsity of an image. In addition, similarity among pixels is integrated into the autoregressive model to further promote the capability and thus an adaptive sparse representation facilitated by a weighted autoregressive model is derived. Since the weighted autoregressive model is inherently determined by the unknown image, we propose a joint optimization scheme by iterative SAR imaging and updating of the weighted autoregressive model to solve this problem. Eventually, experimental results demonstrated the validity and generality of the proposed approach.
BibTeX
@inproceedings{shen2015compressive,
  author = {Fangfang Shen and Guanghui Zhao and Guangming Shi and Weisheng Dong and Chenglong Wang and Yi Niu},
  title = {{Compressive SAR Imaging with Joint Sparsity and Local Similarity Exploitation}},
  booktitle = {Sensors},
  year = {2015},
  url = {https://openreview.net/forum?id=YN9bd3Q9O0},
}
Figure