Generation of reference images using filtered radon transform and truncated SVD for structural artifacts

Seungwan Jeon, Yukyung Lee, Kundong Kim, Daeil Yu, Sung-Su Kim, Joonseo YimIQSP 2022 · DOI

Abstract

Researchers at Samsung Electronics have developed a robust Image Quality Assessment (IQA) method that utilizes Truncated Singular Value Decomposition (SVD) to generate "pseudo-reference" images from distorted chessboard patterns. By combining this with a Radon transform-based alignment, the method precisely quantifies structural artifacts like "zipper edges" and broken lines. The proposed metric achieves a 0.97 correlation with human perception and demonstrates high robustness across varying capture conditions.

Introduction

In modern digital photography, the Image Signal Processor (ISP) is responsible for converting raw sensor data into final images. With the advent of complex color filter arrays, such as Quad Bayer sensors, ISPs must perform sophisticated demosaicing and reconstruction. When ISP tuning parameters are suboptimal, "structural artifacts" emerge—morphological distortions where straight lines appear as "zippers" or fine details appear broken.

Traditional metrics like Signal-to-Noise Ratio (SNR) or standard resolution tests often fail to capture these specific defects. Furthermore, high-quality full-reference IQA typically requires a perfect "ground truth" image, which is difficult to obtain in real-world lab settings due to lighting variations and mechanical misalignments. This paper proposes a method to generate a reliable pseudo-reference image directly from the distorted capture.

Methodology

The proposed evaluation pipeline consists of three primary stages: alignment, pseudo-reference generation, and difference calculation.

Figure 1: Overview of the proposed IQA pipeline, from original image alignment to SSIM-based artifact quantification.

1. Radon Transform-based Alignment

To ensure the mathematical assumptions of the reconstruction hold, the chessboard chart must be perfectly aligned. The system uses a Radon transform—which maps spatial points to lines in a transform domain—to detect the chart's boundaries. This approach is significantly more robust than standard Hough transforms or Harris corner detection, as it operates effectively across varying lighting conditions without requiring manual thresholding.

2. Pseudo-Reference Generation via Truncated SVD

The core innovation lies in exploiting the low-rank property of orthogonal patterns. An ideal chessboard pattern is theoretically a rank-2 matrix. By applying Singular Value Decomposition (SVD) to the warped image and retaining only the top three singular values (Rank 3), the algorithm reconstructs a "clean" version of the image. This process effectively filters out high-rank structural artifacts while preserving the underlying grid geometry and sharpness.

Figure 6: Comparison between (a) the original warped image containing structural artifacts (red arrows) and (b) the reconstructed pseudo-reference image generated via truncated SVD.

3. Difference Calculation

The original distorted image is compared against the SVD-generated pseudo-reference using the Structural Similarity Index (SSIM). This comparison yields a final quality score calibrated to a 0–100 scale.

Experimental Results

Correlation with Human Perception

To validate the metric, a study was conducted with 14 participants providing Mean Opinion Scores (MOS) for various test images. The proposed metric achieved a Pearson Correlation Coefficient (PCC) of 0.97, indicating a near-perfect alignment with human qualitative judgment of structural defects.

Robustness and Sensitivity

In repetition tests under extreme conditions, the method showed a remarkably low standard deviation (0.11). A key finding is the metric's Point Spread Function (PSF) invariance: because the SVD compression maintains the original image's sharpness in the pseudo-reference, the final score isolates structural integrity from lens blur. This allows engineers to distinguish between ISP tuning issues and optical limitations.

Conclusion

This method provides a reliable, automated way to quantify ISP-induced structural artifacts without the need for a physical reference image. While currently optimized for chessboard patterns, the use of truncated SVD offers a powerful framework for isolating systematic noise from geometric structures. Future work will explore the application of this technique to automated ISP optimization loops and alternative chart geometries.

Summary updated

BibTeX
@inproceedings{jeon2022generation,
  author = {Seungwan Jeon and Yukyung Lee and Kundong Kim and Daeil Yu and Sung-Su Kim and Joonseo Yim},
  title = {{Generation of reference images using filtered radon transform and truncated SVD for structural artifacts}},
  booktitle = {IQSP},
  year = {2022},
  doi = {10.2352/ei.2022.34.9.iqsp-341},
}
Figure