Multiscale Edge Detection via Normal Changes

Chwen-Jye Sze, Hong-Yuan Mark Liao, Hai-Lung Hung, Kuo-Chin Fan, Jun-Wei HsiehICIAP 1997 · OpenReview

Abstract

A new edge detection technique based on detection of normal changes is proposed. Most of the existing range image-based edge detection algorithms base their detection criterion on depth or curvature changes. However, the depth change-based approach does not have keen sensitivity in detecting roof (or crease) edges, and the curvature change-based approach suffers from a complicated and tedious principal curvature derivation process. Using normal changes as a detecting criterion, on the other hand, the existence of an edge can be easily detected, even when the change across a boundary is slight. Experimental results using both synthetic and real images demonstrate that the proposed method can efficiently detect both step and roof edges.

Introduction

In computer vision, range images—3D data where each pixel represents a distance—contain two primary types of edges: step edges, characterized by a sudden jump in depth, and roof edges, where two surfaces meet at an angle (crease lines).

Traditional edge detectors often struggle to balance sensitivity and efficiency. Depth-based methods excel at finding step edges but frequently miss subtle roof edges. Conversely, curvature-based methods can identify roof edges but involve computationally expensive and noise-sensitive mathematical derivations. This paper proposes using normal changes—the rate at which the surface's normal vector shifts—as a robust and efficient middle ground for detecting all edge types.

Methodology

Normal Estimation

The first stage of the pipeline involves estimating surface normals from discrete range data. Since derivatives cannot be calculated directly on a grid, the authors evaluate three approximation methods: quadratic surface fitting, orthogonal wavelet-based approximation, and non-orthogonal wavelet-based approximation. The latter is preferred for its superior localization and reduced noise sensitivity.

Geometric framework for normal estimation, showing the surface S, normal vector N, and the Gauss map.

Multiscale Detection

Once normals are estimated, a dyadic wavelet transform (Mallat approach) is applied to analyze normal changes across multiple scales ($2^1, 2^2, 2^3$). Edge points are identified as local extrema where the magnitude of the normal change exceeds a predefined threshold.

The vector of normal change (dN) and its relationship to principal curvatures (κ1, κ2) and principal directions (e1, e2).

The core insight is that the vector of normal change ($dN$) effectively replaces principal curvatures in the detection process, offering comparable descriptive power with significantly reduced mathematical complexity.

Experimental Evaluation

The algorithm was evaluated using both synthetic and real-world datasets:

  • Synthetic: The "agpart" image (240x240), containing various geometric intersections.
  • Real-world: The "occl" and "bigwye" range images, featuring complex overlapping objects and diverse surface orientations.

Key Insights and Results

  • The "Goldilocks" Metric: Normal changes are more sensitive than depth changes but less computationally intensive than curvature. The magnitude of the normal change is independent of the coordinate basis, making it a highly efficient proxy for surface boundaries.
  • Wavelet Superiority: Non-orthogonal wavelet-based approximation outperformed other methods, which often "delocalized" edges or introduced spurious results.
  • Crease Detection: The method successfully identified "crease" (roof) edges that are typically invisible to standard depth-based detectors.
  • Scale Robustness: The multiscale approach allowed the detector to distinguish sharp features from noise, yielding cleaner edge maps at higher scales ($2^2$ and $2^3$).

Conclusion

While the method relies on preset thresholds and requires a two-stage process, it offers a significant advancement in range image edge detection. By leveraging normal changes and non-orthogonal wavelets, it achieves a balance of sensitivity and efficiency that traditional methods lack. Future work may focus on refining the wavelet approximation to further enhance edge localization in noisy environments.

Summary updated

Abstract

A new edge detection technique based on detection of normal changes is proposed. Most of the existing range image-based edge detection algorithms base their detection criterion on depth or curvature changes. However, the depth change-based approach does not have keen sensitivity in detecting roof ( or crease ) edges, and the curvature change-based approach suffers from a complicated and tedious principal curvature derivation process. Using normal changes as a detecting criterion, on the other hand, the existence of an edge can be easily detected, even when the change across a boundary is slight. Experimental results using both synthetic and real images demonstrate that the proposed method can efficiently detect both step and roof edges.
BibTeX
@inproceedings{sze1997multiscale,
  author = {Chwen-Jye Sze and Hong-Yuan Mark Liao and Hai-Lung Hung and Kuo-Chin Fan and Jun-Wei Hsieh},
  title = {{Multiscale Edge Detection via Normal Changes}},
  booktitle = {ICIAP},
  year = {1997},
  url = {https://openreview.net/forum?id=BASRXJMtKd},
}
Figure