Calibration de la métrique MS-SSIM pour les distorsions de compression à l'aide d'une échelle perceptive des différences

Christophe Charrier, Kenneth Knoblauch, Laurence T. Maloney, Alan C. BovikHAL 2011

TL;DR

This paper tackles the challenge of accurately assessing image compression quality by recalibrating the widely-used MS-SSIM metric. By using a novel perceptual difference scaling method (MLDS) and an optimization technique, they demonstrate a significant improvement in MS-SSIM's correlation with human perception, particularly for compression artifacts.

What Problem Does This Paper Address?

Evaluating the quality of compressed digital images is crucial across many industries. While metrics like MS-SSIM (Multi-Scale Structural Similarity) offer a good correlation with human judgment (MOS - Mean Opinion Score), they often rely on parameters that are not optimally tuned for specific types of image degradation, like compression. This can lead to discrepancies, especially when predicting subtle perceptual differences between images with varying compression levels. The authors aim to bridge this gap by developing a method to fine-tune MS-SSIM's parameters to better align with human perceptual assessments of compression artifacts.

Methods

The core of this paper lies in two key methodological contributions:

  • Maximum Likelihood Difference Scaling (MLDS): This technique is used to quantify supra-threshold perceptual differences between pairs of images. It allows researchers to build a perceptual scale of how "different" two images appear to human observers, regardless of the compression scheme. MLDS is based on the idea that if observer responses to image pairs (a:b) and (c:d) indicate that (a:b) is a larger difference than (c:d), then the underlying perceptual distances between a and b are greater than between c and d.
  • Genetic Algorithm Optimization: To recalibrate MS-SSIM, the authors employ a genetic algorithm. This algorithm searches for the optimal values of MS-SSIM's 15 parameters (weights for luminance, contrast, and structure across different scales) by minimizing an error function. This error function quantifies the difference between the MLDS-derived perceptual scores and the MS-SSIM predictions.

Datasets

The paper utilizes two established image quality assessment databases for evaluation:

  • LIVE Image Quality Assessment Database (v2): This database contains images with various types of distortions, including compression.
  • TID2008 Database: This database also features a wide range of distortions, including those relevant to compression.

The authors also mention using a specific set of 15 original JPEG2000 compressed images at varying bitrates (0.1000 to 1.5912 bpp) for their MLDS experiments.

Key Insights and Ideas

  • MLDS as a Ground Truth for Compression: The MLDS method proves to be a valuable tool for creating a perceptually accurate scale of image quality degradation due to compression, offering a more nuanced view than traditional MOS scores alone.
  • Parameter Optimization is Key: The paper strongly suggests that the default parameters in MS-SSIM are not universally optimal. Tailoring these parameters, especially for the structural distortion component, can significantly improve its predictive power for compression artifacts.
  • Structural Distortion Matters Most: The study highlights that the structural distortion factor in MS-SSIM is particularly sensitive to compression artifacts. Optimizing its associated parameters yields the most substantial gains in correlation with human perception.
  • Luminance and Contrast are More Uniformly Weighted: For luminance distortion, the optimization process reveals that all five decomposition levels contribute almost equally, suggesting a more balanced role for these factors.

Key Results

  • The recalibrated MS-SSIM metric, using the optimized parameters, shows a significantly improved level of correspondence with the MLDS perceptual scales compared to the original MS-SSIM. This is visually demonstrated in Figure 1, where the black dotted curve (optimized MS-SSIM) closely follows the black dots (MLDS values).
  • Table 2 and Table 3 present correlation coefficients (Pearson, Kendall, and Spearman) that demonstrate the improvement. For specific compression-related degradations (e.g., JPEG2000, JPEG), the optimized MS-SSIM shows statistically significant gains in correlation with human judgments.
  • While the overall improvement is notable, the paper acknowledges that the gains are not as significant for certain noise-based degradations like white noise or Gaussian blur.

Limitations and Future Work

  • Limited Scope of Optimization: The optimization was performed on a specific set of 15 JPEG2000 compressed images. While this provided valuable insights, extending the optimization to a broader range of compression types and image content could yield even more robust results.
  • Non-convex Error Function: The authors acknowledge that the error function is non-convex, meaning genetic algorithms are crucial for finding a global minimum, but there's always a possibility of not reaching the absolute best set of parameters.
  • Future Work: The authors suggest that further research could explore other optimization strategies and investigate the impact of these optimized parameters on different image quality assessment tasks beyond compression. They also imply that the MLDS framework itself could be further refined for even more precise perceptual scaling.

Summary updated

Abstract

National audience
BibTeX
@inproceedings{charrier2011calibration,
  author = {Christophe Charrier and Kenneth Knoblauch and Laurence T. Maloney and Alan C. Bovik},
  title = {{Calibration de la métrique MS-SSIM pour les distorsions de compression à l'aide d'une échelle perceptive des différences}},
  booktitle = {HAL},
  year = {2011},
}
Figure