Correlation Preserving Sparse Coding Over Multi-level Dictionaries for Image Denoising

Rui Chen, Huizhu Jia, Xiaodong Xie, Wen GaoarXiv 2016 · OpenReview

TL;DR

This paper introduces a novel image denoising method called Correlation Preserving Sparse Coding (CPSC). It tackles the issue of lost correlations in noisy images during sparse coding by incorporating two novel regularization strategies: an Adaptive Graph Laplacian (AGL) for global similarity preservation and Locality-Constrained (LC) coding for local feature consistency. The results show state-of-the-art denoising performance, outperforming existing methods.

What Problem Does This Paper Address?

Image denoising aims to remove noise from an image while preserving its essential details like edges and textures. Traditional methods often struggle because the noise corrupts the underlying structure and correlations within the image. Dictionary learning-based methods, which represent image patches as sparse combinations of dictionary elements, are powerful but can suffer from a loss of crucial correlations. This happens because the process of converting high-dimensional image patches into vectors can discard local spatial relationships, and learning dictionaries from noisy patches can lead to unstable representations. The paper's core problem is how to effectively preserve both global and local correlations within image patches during the sparse coding process to achieve superior denoising.

Methods

The proposed CPSC method introduces two key regularization strategies to preserve correlations:

  • Adaptive Graph Laplacian (AGL) Regularization: This component leverages spectral graph theory to preserve global similarity correlations. It constructs a graph where data points are vertices and edge weights represent similarity. By using a normalized graph Laplacian, it aims to capture the intrinsic geometric structure of the data. The edge weights are derived from a non-negative low-rank construction, which helps in robustly identifying global data structures.
  • Locality-Constrained (LC) Coding: This strategy focuses on preserving local feature correlations. It ensures that similar image patches are represented by similar sparse codes. This is achieved by using a locality adaptor that weights dictionary elements based on their similarity to the input patch. This helps maintain internal consistency within noisy patches and prevents the loss of local discriminative features.

These two strategies are integrated into a unified CPSC model, which is then optimized using an alternating iterative strategy to update the dictionary and sparse coefficients.

Datasets

The paper evaluates its method on seven natural images. Specific images mentioned include "Barbara," "Boat," "Monarch," "Wheel," "Airplane," "Lena," and "House." The noisy versions of these images were simulated by adding white Gaussian noise with varying standard deviation levels (20 to 80). No new datasets were introduced.

Key Insights and Ideas

  • Correlation is King: The paper strongly emphasizes that preserving correlations (both global and local) is paramount for effective image denoising, especially when using dictionary learning.
  • AGL for Global Structure: The AGL regularizer, by incorporating low-rank construction for edge weights, effectively captures and preserves the overarching geometric structures of image patches, which is crucial for distinguishing between signal and noise.
  • LC for Local Detail: The LC coding ensures that fine-grained local features and neighborhood consistencies are maintained, preventing the loss of texture and edge information that can occur with standard sparse coding.
  • Unified Approach: Combining AGL and LC within a single framework allows for a more comprehensive preservation of image structure, leading to better denoising performance than methods that focus on only one type of correlation.
  • Robustness to Noise: The method demonstrates improved performance, particularly at higher noise levels, suggesting its ability to disentangle signal from noise more effectively.

Key Results

  • State-of-the-Art Performance: The CPSC method achieved state-of-the-art denoising performance, outperforming established methods like BM3D, EPLL, and NSCR in terms of both Peak Signal-to-Noise Ratio (PSNR) and subjective visual quality.
  • Quantitative Improvements: Table I shows that the proposed method consistently achieved higher PSNR values across all tested images and noise levels. For example, on the "Barbara" image with a noise level of 60, CPSC achieved a PSNR of 31.04 dB, compared to 29.87 dB for NSCR, 29.45 dB for EPLL, and 29.28 dB for BM3D. The average PSNR across all images was 29.74 dB, significantly higher than the other methods.
  • Qualitative Improvements: Figures 1 and 2 visually demonstrate the superior denoising capabilities. The proposed method effectively removes noise while preserving sharper edges and smoother regions, resulting in more visually pleasing and artifact-free images, especially under high noise conditions.

Limitations and Future Work

  • Parameter Tuning: The paper mentions that regularization parameters ($\mu$, $\rho$, $\lambda$) and other parameters ($\alpha$, $\beta$, $\delta$) need to be empirically tuned. This suggests that achieving optimal performance might require careful parameter selection for different types of images or noise.
  • Computational Complexity: While the paper mentions fast convergence (within 100 iterations), dictionary learning methods can still be computationally intensive. The paper doesn't explicitly detail the computational cost compared to other methods.
  • Future Work: The authors suggest that their method's ability to preserve intrinsic correlations makes it competitive and potentially applicable to other image restoration tasks beyond denoising. Further exploration into these applications could be a future direction.

Summary updated

Abstract

In this letter, we propose a novel image denoising method based on correlation preserving sparse coding. Because the instable and unreliable correlations among basis set can limit the performance of the dictionary-driven denoising methods, two effective regularized strategies are employed in the coding process. Specifically, a graph-based regularizer is built for preserving the global similarity correlations, which can adaptively capture both the geometrical structures and discriminative features of textured patches. In particular, edge weights in the graph are obtained by seeking a nonnegative low-rank construction. Besides, a robust locality-constrained coding can automatically preserve not only spatial neighborhood information but also internal consistency present in noisy patches while learning overcomplete dictionary. Experimental results demonstrate that our proposed method achieves state-of-the-art denoising performance in terms of both PSNR and subjective visual quality.
BibTeX
@misc{chen2016correlation,
  author = {Rui Chen and Huizhu Jia and Xiaodong Xie and Wen Gao},
  title = {{Correlation Preserving Sparse Coding Over Multi-level Dictionaries for Image Denoising}},
  year = {2016},
  eprint = {1612.08049},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1612.08049},
}
Figure