A Multiscale and Anisotropic Edge Detection Algorithm

Hua-Jie Cai, Xin Tian, Tao LiAdvances 2014 · DOI

Abstract

Edge detection is of fundamental importance in computer vision. Current methods based on first or second derivatives, such as Canny, Laplace, and Gaussian operators, typically rely on single-scale information and do not sufficiently account for edge orientation. Consequently, these operators often struggle to distinguish true edges from noise. This paper proposes a multiscale and multidirectional edge detection algorithm for grayscale images. Using a Gaussian function as the filter kernel, a series of multiscale and directional filters are generated. A bank of edge maps is acquired by convolving the original image with these filters, and the maximum response is used to identify local maxima. Finally, edges are determined through thresholding. Experimental results demonstrate that the proposed algorithm outperforms several state-of-the-art methods.

Introduction

Edge detection serves as a critical building block for higher-level computer vision tasks. However, classical operators like Canny, Sobel, and the Laplacian of Gaussian (LoG) face persistent challenges. Most traditional operators are isotropic or limited to simple horizontal and vertical gradients, making them less effective at resolving diagonal edges. Furthermore, they face a "scale trade-off": small-scale filters capture fine details but are sensitive to noise, while large-scale filters suppress noise at the cost of blurring or shifting edge locations. Standard Canny detection is particularly sensitive to impulsive "salt and pepper" noise, as its initial smoothing step can mistakenly amplify noise spikes into edge features.

Methodology

The proposed algorithm leverages Scale Space theory and directional filtering to overcome these traditional hurdles through a multi-step pipeline:

  1. Anisotropic FDOG Filter Bank: Instead of symmetric filters, the researchers generate a bank of First Derivative of Gaussian (FDOG) filters. These are anisotropic (elongated), making them highly sensitive to specific orientations.
  2. Multidirectional Rotation: Filters are rotated into six specific directions ($\theta = \{0, \pi/6, \pi/3, \pi/2, 2\pi/3, 5\pi/6\}$) across four different scales, resulting in 24 unique filters.
  3. Maximum Response Extraction: The image is convolved with this filter bank. For every pixel, the algorithm selects the maximum response across all directions to ensure edges are captured at their strongest regardless of orientation.
  4. Scale Multiplication: To distinguish true structural edges from random noise, the authors multiply the response maps of adjacent scales. Since true edges persist across scales while noise does not, this step amplifies the signal while suppressing artifacts.
  5. Thresholding: A final threshold is applied to the product of the scale multiplication to produce the binary edge map.
Figure 1: Flowchart of the proposed multiscale and multidirectional edge detection pipeline.

Experimental Results

The algorithm was evaluated using standard test images (such as "cameraman.bmp") and synthetic images corrupted with Gaussian and impulsive noise.

In low-noise environments, the proposed method performs comparably to the Canny operator but provides superior detail in complex regions. The algorithm's true strength lies in its high-noise robustness. In tests involving heavy degradation, traditional Sobel and Canny operators either failed completely or produced significant false positives. By utilizing directional information—which noise lacks—and scale multiplication, the proposed method maintains clean structural contours and high localization accuracy even in heavily degraded images.

Figure 2: Comparative edge detection results showing the clarity and continuity of the extracted features.

Conclusion

This multiscale and anisotropic approach significantly improves edge detection by moving beyond the horizontal/vertical constraints of traditional methods. While the convolution with 24 filters increases computational complexity compared to standard Canny detection, the gains in noise suppression and edge localization are substantial. Future work will focus on optimizing the algorithm for real-time applications via hardware acceleration and developing adaptive thresholding mechanisms to improve autonomy across diverse imaging environments.

Summary updated

Abstract

Edge detection is of important significance for computer vision. Current edge detection methods based on the first derivative or second derivative such as Canny and Laplace and Gaussian operators use single scale information and do not take account of the edge directions sufficiently. Those operators cannot distinguish edges and noise well. A multiscale and multidirectional edge detection algorithm is proposed for gray images in this paper. Gaussian function is used as the filter kernel. A serials scales and directional filters are generated by the generating function. A bank of edge maps are acquired by convoluting the original image with those filters. The maximum response is used to find the local maximum in the maps. Finally, the edges are determined after a certain threshold. The experiments show that the proposed algorithm gets outperformance compare to some state of art methods.
BibTeX
@inproceedings{cai2014multiscale,
  author = {Hua-Jie Cai and Xin Tian and Tao Li},
  title = {{A Multiscale and Anisotropic Edge Detection Algorithm}},
  booktitle = {Advances},
  year = {2014},
  doi = {10.2991/lemcs-14.2014.249},
}
Figure