This paper establishes a theoretical and quantitative foundation for identifying salient features in images by comparing filter responses to a normal distribution. By framing salience as a "suspicious coincidence" that deviates from a random baseline, the authors demonstrate that orientation energy in natural images follows a power law, while random noise remains near-Gaussian. This distinction allows for highly accurate, objective edge detection.
Introduction
In computer vision, while basic edge detection is well-understood, determining "salience"—which edges are actually prominent or meaningful—remains a significant challenge. Humans effortlessly distinguish meaningful contours from background texture or noise, but replicating this computationally has historically lacked objective validation and theoretical justification. Previous research noted that orientation filter responses in natural images follow a "heavy-tailed" non-Gaussian distribution. By comparing this to a normal distribution of the same variance, researchers could identify unusually high responses as salient. This study formalizes why the normal distribution serves as the ideal baseline for non-salience.
Methodology
The authors utilize oriented Gabor filters, which mimic the receptive fields of neurons in the primary visual cortex. The processing pipeline involves:
- Preprocessing: Applying a Difference of Gaussian (DoG) filter.
- Filtering: Convolving the image with Gabor filters at six orientations.
- Energy Calculation: Computing the "Orientation Energy" (E) for each pixel.
- Thresholding: Comparing the actual orientation energy distribution $h(E)$ to a matching normal distribution $g(E)$ with the same variance.
A key innovation in this work is the use of synthetic images—overlapping squares with varying levels of embedded uniform noise (1%, 5%, and 10%)—to provide a "ground truth" for salience. This allows for the calculation of the Sum of Squared Error (SS Error) between thresholded noisy images and clean reference images.
Theoretical Framework: The Gaussian Baseline
The core of the authors' argument is that salience is a deviation from an "unsuspicious" baseline. If an image were purely random white noise, its filter responses would be independent and follow a near-Gaussian distribution. This provides the mathematical justification for using the normal distribution: it is the statistical signature of "nothing interesting happening."
As shown in the figure above, natural images follow a power-law distribution ($1/x^a$). The point where the power law begins to significantly exceed the Gaussian (the "heavy tail") is the mathematically optimal threshold for salience, aligning with Horace Barlow’s theory of "suspicious coincidence."
Key Results
The experiments demonstrated that Orientation Energy Distribution (OED) thresholding overwhelmingly outperforms fixed-percentile methods. In tests using synthetic images with 5% noise, the difference in SS Error was statistically significant ($p < 0.025$).
Furthermore, the authors developed a thresholding method based directly on white-noise statistics that achieved a 0.98 correlation with human-selected salience thresholds, a marked improvement over the 0.91 correlation found in previous studies. This suggests that the human visual system may use similar statistical comparisons to identify salient features.
Conclusion
The "suspicious coincidence" framework provides a robust method for objective salience detection by comparing observed distributions to a random baseline. While this study focused on orientation energy, the authors suggest this approach could be extended to other sensory tasks and image features, such as color, motion, or spectral histograms used in texture segmentation. Any peaked distribution with high kurtosis could potentially be analyzed using this "unsuspicious baseline" methodology.
Summary updated