Canny Edge Detection

The low and high thresholds control the hysteresis procedure that filters detected edges. Pixels with gradient strength above the high threshold (default 200) are always kept, while those between the low threshold (default 100) and high threshold are preserved only if connected to strong edges. Lowering the low threshold (10–30) captures fine details and subtle contours, whereas raising it (150–200) strips out internal textures to leave only bold silhouettes. OpenCV recommends maintaining a 1:2 or 1:3 ratio between low and high thresholds for optimal connectivity.

Use Canny edge detection when you need precise geometric alignment, razor-sharp silhouettes, or exact line contours such as anatomical poses, architectural profiles, logos, and hard-surface products. Because Canny creates a 2D binary boundary map, it excels at maintaining exact physical borders. However, if your scene relies on volumetric spatial depth, 3D surface curvature, or soft organic forms, depth maps or normal maps are better suited because Canny cannot perceive 3D volume or subtle tonal gradients.

High-contrast, sharp images with clear subject-background separation produce the cleanest edge maps. In contrast, blurry, low-resolution, or heavily compressed JPEG images generate fragmented, noisy lines. Similarly, foggy scenes or subtle low-contrast gradients may cause the algorithm to miss edges entirely, while intricate micro-textures like dense foliage or knit wool can produce chaotic speckle noise.

First developed by John F. Canny in 1986, the algorithm operates through a four-stage classical computer vision pipeline. It first applies a Gaussian filter to smooth image noise, then calculates gradient intensity and direction using a Sobel operator. Next, non-maximum suppression thins detected edge boundaries down to crisp single-pixel widths. Finally, dual-threshold hysteresis filters out weak noise while preserving continuous structural contours.