Canny Edge Detector in Practice (circa 2013)

· computer-vision, image-processing

Canny’s blend of Gaussian smoothing, gradient estimation, non-maximum suppression, and hysteresis linking made it the go-to detector for general-purpose pipelines in 2013. Below is a practical cheat-sheet for replicable results on CPUs and modest GPUs from that period.

Parameter recipes

  • Gaussian sigma (σ): 1.0–1.4 for VGA/720p inputs; 1.6–2.0 for 1080p when you want broader edges and better noise suppression.
  • High/low thresholds: A common ratio was 2:1 or 3:1. Many OpenCV users started around high = 80–120 and low = 40–60 for 8-bit images, adjusting upward for sharper textures or downward for smoother scenes.
  • Gradient operator: Sobel (3×3) was the default for its noise robustness; Scharr (3×3) delivered slightly better rotation fidelity and was favored when texture orientation mattered.

Practical tuning workflow

  1. Normalize lighting first: Apply a mild bilateral filter or CLAHE before Canny when illumination is uneven.
  2. Pick sigma, then thresholds: Choose σ to match the expected edge scale; only then sweep thresholds to balance recall vs. false positives.
  3. Use non-maximum suppression diagnostics: Visualize gradient angles to verify that edges are being thinned rather than broken.
  4. Guard against speckle: If salt-and-pepper noise is present, precede Canny with a 3×3 median filter.

Performance notes (2013 hardware)

  • SIMD mattered: SSE2/AVX intrinsics in OpenCV 2.x gave 2–4× speedups over naive loops.
  • GPU ports were early-stage: CUDA implementations existed but incurred PCIe overhead; they paid off mainly for batch processing or >1080p frames.
  • Edge density considerations: High-texture scenes could flood hysteresis; clamping the high threshold or applying a light Gaussian blur first kept runtime predictable.

When to use LoG instead

If edge polarity is ambiguous or you need rotational symmetry without gradient direction bookkeeping, Laplacian of Gaussian (LoG) can still be competitive. In 2013-era tests, Canny generally won on localization and thin edges, while LoG produced slightly smoother, more continuous contours on very noisy inputs.

Example OpenCV 2.x call

cv::Mat edges;
cv::GaussianBlur(img, img, cv::Size(0, 0), 1.4);
cv::Canny(img, edges, 50, 120, 3, true); // thresholds, aperture size, use L2 gradient

Keep the thresholds proportional when changing σ or aperture size. For production, profile with and without the extra GaussianBlur—OpenCV’s internal smoothing can suffice for many scenes.

Keep Reading