An optimized direction statistics for detecting and removing random-valued impulse noise

Hao Yang, Hang Qiu, Leiting ChenJournal of Industrial and Management Optimization 2017 · DOI

TL;DR

This paper introduces WROD-EPR, a novel two-stage method for effectively detecting and removing random-valued impulse noise from images. It achieves superior performance by employing an optimized direction-based statistic (WROD) for accurate noise identification, especially around edges, and then using an edge-preserving regularization technique for robust noise restoration.

What Problem Does This Paper Address?

Digital images are often degraded by impulse noise, which randomly replaces pixel values. This noise can severely damage image information, hindering subsequent processing tasks like segmentation and object recognition. While various de-noising methods exist, they often struggle to preserve image details and edges, especially at high noise levels. Existing methods that attempt to detect noise before filtering can still blur fine features or misclassify edge pixels as noise. This paper aims to address this by developing a more accurate noise detection mechanism that specifically accounts for image structures like edges, leading to better overall image restoration.

Methods

The proposed WROD-EPR method operates in two stages:

  1. Noise Detection (WROD Statistic): This stage introduces a new local image statistic called WROD (Weight Rank-Ordered Difference). It builds upon the existing ROAD (Rank-Ordered Absolute Differences) statistic but enhances it by:
  2. Optimized Direction Statistics: Instead of just considering all neighbors, WROD analyzes pixel differences along specific directions (horizontal, vertical, and two diagonals) within a larger sliding window (typically 5x5).
  3. Weighted Standard Deviation: It calculates the standard deviation of pixel intensities along these directions, weighting them based on their distance from the center pixel. This helps to better capture local image structure.
  4. Minimum Standard Deviation Direction: The direction with the minimum standard deviation is identified, indicating the most coherent local structure (likely an edge or a smooth region).
  5. Extensional Rank Candidate Set (Ep): A set of candidate pixels is formed based on the minimum standard deviation direction, focusing on pixels that are likely to be noise-free if the center pixel is part of an edge.
  6. WROD Calculation: The WROD statistic is then computed using the absolute differences of pixels within this candidate set, making it more robust to noise, especially at edges.
  7. Noise Restoration (Edge-Preserving Regularization): Once noisy pixels are identified by the WROD detector, they are restored using an edge-preserving regularization (EPR) technique. This stage focuses solely on the regularization term, omitting a data-fitting term because the non-noisy pixels are assumed to be correctly represented by their original values. The method uses an iterative approach with a decreasing threshold for noise detection to progressively refine the set of noisy pixels.

Datasets

The paper evaluates its method on standard 8-bit grayscale images, including "Lena," "Bridge," and "Pentagon." These images are corrupted with random-valued impulse noise at various densities (40%, 50%, 60%, 70%). No new datasets or benchmarks are introduced.

Key Insights and Ideas

  • Edges are Crucial: The core insight is that accurately detecting impulse noise, especially around edges, requires considering local image structure beyond simple pixel intensity differences. The WROD statistic's focus on optimized directions and weighted standard deviations directly addresses this.
  • Direction Matters: By analyzing pixel relationships along specific directions, the method can better distinguish between true image edges and noise-induced intensity variations.
  • Two-Stage Approach Synergy: The combination of a robust noise detector (WROD) with an effective edge-preserving restorer (EPR) is key. The accuracy of the detector directly impacts the quality of the restoration.
  • Robustness at High Noise Levels: The WROD statistic's design allows it to maintain accuracy even when the noise density is very high (up to 60%), a common challenge for many existing methods.

Key Results

  • Superior Noise Detection: The proposed WROD detector demonstrates a better balance between detecting true noisy pixels ("miss") and incorrectly flagging clean pixels ("false-hit") compared to several other methods, especially at higher noise ratios.
  • Enhanced Image Restoration: The WROD-EPR filter achieves higher Peak Signal-to-Noise Ratio (PSNR) values across various images and noise densities compared to established methods like ACWM, Luo, ROR-NLM, and ROLD-EPR. For instance, on the "Lena" image at 60% noise, the proposed method achieves a PSNR of 29.56, outperforming others.
  • Effective Edge and Detail Preservation: Qualitative results show that the WROD-EPR method preserves image edges and fine details significantly better than competing algorithms, avoiding the blurring artifacts often seen in other approaches.
  • Computational Efficiency: While the detection stage is more complex than simpler methods, the overall time cost is competitive, especially at higher noise densities where its detection accuracy leads to more efficient restoration. The detection time remains relatively constant regardless of noise density.

Limitations and Future Work

  • Computational Cost: While the method is efficient, the detection stage is noted to be more computationally intensive than simpler filters, though the paper argues this cost is acceptable for the improved accuracy.
  • Parameter Tuning: The paper mentions the need to tune parameters like a in the weighted direction index and the initial threshold T0 for the WROD detector. While they provide optimal values based on experiments, these might require adjustment for different image types or noise characteristics.
  • Future Work: The paper doesn't explicitly detail future work, but the success of the WROD statistic suggests potential for its application in other image processing tasks that require robust local feature extraction and noise-robustness. Further exploration of adaptive thresholding strategies or more sophisticated regularization terms could also be avenues for future research.

Summary updated

Abstract

In this paper, we propose a robust local image statistic based on optimized direction, by which we can distinguish image details and edges from impulse noise effectively. Therefore it can identify noisy pixels more accurately. Meanwhile, we combine it with the edge-preserving regularization to remove random-valued impulse noise in the cause of precise estimated value. Simulation results show that our method can preserve edges and details efficiently even at high noise levels.
BibTeX
@article{yang2017optimized,
  author = {Hao Yang and Hang Qiu and Leiting Chen},
  title = {{An optimized direction statistics for detecting and removing random-valued impulse noise}},
  journal = {Journal of Industrial and Management Optimization},
  year = {2017},
  doi = {10.3934/jimo.2017062},
}
Figure