Abstract
Standard edge detectors often fail because they assume all edges are simple "steps." This paper proves that linear filters (like the Canny detector) inherently struggle to accurately locate "composite" edges—mixtures of steps, peaks, and roofs—and introduces a quadratic filtering approach that achieves zero localization error while remaining immune to false positives caused by smooth shading.
Introduction: The Problem with Step Edges
In the early 1990s, computer vision relied heavily on the "step edge" model: the idea that an edge is a sudden jump in brightness. However, real-world physics is messier. Due to mutual illumination, specularities, and surface geometry, edges are often a combination of steps, delta functions (lines), and roofs (ramps).
The authors identify two major flaws in existing linear edge detectors:
- Systematic Localization Error: When an edge is "composite" (e.g., a step plus a small peak), a linear filter will shift the detected edge position away from the true boundary.
- Shading Sensitivity: Linear filters often mistake smooth, constant brightness gradients (shading) for actual edges, leading to "false positives" unless high thresholds are used, which might then discard real, low-contrast edges.
Methods: Moving to Quadratic Filtering
The core innovation is moving from linear filtering to quadratic filtering. Instead of looking for maxima in a single linear filter response, the authors look for maxima in the "energy" of the signal. This is calculated as the sum of squares of responses from a family of filters: $W = \sum (I * f_i)^2$.
Specifically, they use Hilbert pairs—pairs of even and odd filters (like a second derivative of a Gaussian and its Hilbert transform). This makes the detector "phase-independent," allowing it to detect a step, a line, or a roof at the exact same location without bias. Furthermore, they show that the Hilbert transform of the second derivative of a Gaussian ($G''_H$) is particularly effective, offering 32% better immunity to shading than standard Gaussian derivatives.
To handle 2D images, a windowed approach and a technique based on Singular Value Decomposition (SVD) are used to approximate the filter response at any orientation. This allows the system to represent multiple orientations at a single point, preventing the "rounding" of corners and junctions common in other detectors.
Key Results and Localization Accuracy
The authors provide a mathematical proof that any fixed, finite family of linear filters will have systematic localization errors for certain composite edges. In contrast, the quadratic filter achieves zero systematic localization error for composite edges.
Evaluation on synthetic 1D signals proved zero localization error for various composite edge profiles. Qualitatively, the detector maintains the integrity of T-junctions where standard detectors often "break" or round the intersection. By allowing for multiple dominant orientations at a single pixel, the algorithm can localize T-junctions and corners without typical blurring or "breaking" seen in detectors that assume only one edge direction per pixel.
Conclusion
While computationally more complex than linear methods at the time of publication, this approach addresses fundamental theoretical limits of edge detection. By moving to a phase-independent quadratic energy model, the system provides a more robust and accurate foundation for higher-level computer vision tasks like surface reconstruction and object recognition.
Summary updated