Method: cv::Canny()
Description
cv::Canny() performs edge detection using the Canny algorithm.
Syntax
cv::Canny(src, edges, threshold1, threshold2, apertureSize);Parameters
src(cv::Mat): Input single-channel (grayscale) image.edges(cv::Mat): Output binary image with detected edges.threshold1(double): Lower threshold for the hysteresis procedure.threshold2(double): Upper threshold for the hysteresis procedure.apertureSize(int, optional): Aperture size for Sobel operator (default: 3).
Parameters Explained
| Parameter | Type | Description |
|---|---|---|
| lowThreshold | double | Lower boundary for hysteresis thresholding. |
| highThreshold | double | Upper boundary for hysteresis thresholding. |
| apertureSize | int | Size of the Sobel operator kernel (optional). |
| L2gradient | bool | Use more accurate L2 norm (optional). |
Hysteresis Thresholding
Hysteresis thresholding uses two thresholds: lowThreshold and highThreshold. Pixels with gradient magnitudes above the high threshold are marked as strong edges, while pixels below the low threshold are discarded. Pixels between the thresholds are considered weak edges and are only included if they are connected to strong edges, ensuring that edges are continuous and accurate.
Result Visualization
Hysteresis Thresholding in Canny Edge Detection
In the context of OpenCV’s Canny function, hysteresis refers to the double-thresholding step that decides which gradient magnitudes correspond to “true” edges. Instead of a single cutoff, Canny uses two thresholds:
-
threshold1(the low threshold) — any pixel with gradient magnitude below this is immediately discarded. -
threshold2(the high threshold) — any pixel with gradient magnitude above this is immediately marked as a “strong” edge.
Why use two thresholds?
A single threshold can either miss faint but continuous edges (if set too high) or include a lot of noise (if set too low). Hysteresis overcomes this by:
- Clearly accepting pixels with very strong gradients.
- Rejecting pixels with very weak gradients.
- Conditionally accepting “weak” pixels whose gradient lies between the two thresholds, but only if they are connected (8-connected neighbourhood) to a strong edge pixel.
OpenCV Canny Signature
// src: input grayscale image (CV_8U or CV_16U)
// dst: output binary edge map (edges drawn in white on black background)
double lowThresh = 50.0; // threshold1
double highThresh = 150.0; // threshold2
int apertureSize = 3; // Sobel kernel size (must be odd: 3, 5, or 7)
bool L2gradient = false; // false: use L1 norm, true: use L2 norm
// Apply Canny with hysteresis thresholding
cv::Canny(src, dst, lowThresh, highThresh, apertureSize, L2gradient);
How it works internally
// 1. Compute gradient magnitude G and direction θ at each pixel.
// 2. Non-maximum suppression: thin edges by keeping only local maxima along θ.
// 3. Double-threshold (hysteresis):
// - If G(x,y) ≥ highThresh → mark as STRONG_EDGE.
// - If G(x,y) < lowThresh → mark as NO_EDGE.
// - If lowThresh ≤ G(x,y) < highThresh → mark as WEAK_EDGE.
// 4. Edge tracking by connectivity:
// - For each WEAK_EDGE pixel, check 8 neighbours.
// - If at least one neighbour is a STRONG_EDGE, promote this pixel to STRONG_EDGE.
// - Otherwise, suppress it (set to NO_EDGE).
// 5. Final binary map: all STRONG_EDGE pixels become white (255), others black (0).
How Different Hysteresis Thresholds Affect Edge Filtering
Recall: Canny uses two thresholds— low (t₁) and high (t₂)—and only keeps “weak” edges (t₁ ≤ gradient < t₂) if they connect to a “strong” edge (gradient ≥ t₂).
1. Very Permissive: t₁=20, t₂=40
- Keeps: Almost every brightness change, including fine textures (hair, fabric weave, grass).
- Filters out: Only the tiniest fluctuations (<20)—but most image noise still appears as edges.
2. Permissive: t₁=50, t₂=100
- Keeps: Clear edges (object outlines, strong shadows) and many medium-strength details (minor surface scratches, fine print).
- Filters out: Low-contrast texture and most sensor noise below gradient 50.
3. Balanced: t₁=50, t₂=150
- Keeps: Major structural edges (walls, tables, faces) and faint edges only if linked to them.
- Filters out: Isolated medium details and moderate-contrast texture.
4. Strict: t₁=100, t₂=200
- Keeps: Only the strongest, most obvious edges (high-contrast borders, sharp object outlines).
- Filters out: Almost all fine detail, texture, and noise below gradient 100.
Example Code Snippet
// Very permissive: captures texture & noise
cv::Canny(src, dst1, 20, 40);
// Permissive: captures clear edges + some details
cv::Canny(src, dst2, 50, 100);
// Balanced: favors true object outlines
cv::Canny(src, dst3, 50, 150);
// Strict: only strongest edges
cv::Canny(src, dst4, 100, 200);
Key takeaway: - Lower thresholds → more edges (including noise & texture). - Higher thresholds → fewer, stronger edges (filtering out faint details). - The gap between t₁ and t₂ (hysteresis) lets you keep meaningful weak edges while discarding isolated noise.
Tuning Hysteresis
-
Low : High ratio — typically between
1:2and1:3. A smaller ratio (e.g.1:2) is more permissive, capturing finer details but risking noise. - Aperture size — affects gradient estimation; larger sizes smooth more, which can suppress noise but blur edges.
-
L2gradient flag — using the true Euclidean norm (
L2gradient=true) gives slightly more accurate edge strength than the default sum of absolute derivatives (L1), at a small performance cost.
By understanding hysteresis thresholding, you can more effectively adjust your two thresholds to balance noise suppression against edge completeness in your application.
Example Usage
cv::Mat edges;
cv::Canny(img, edges, 100, 200);
cv::imshow("Edges", edges);