Walk me through Canny edge detection and why it beats Sobel thresholding

WHAT THIS TESTS: Whether you view edge detection as a structured signal-processing pipeline or a single filter operation. Interviewers want to see that you understand why raw gradient magnitude thresholding produces thick, noisy, and discontinuous boundaries, and how Canny's design addresses each failure mode through explicit stages rather than magic.

A GOOD ANSWER COVERS: Five stages in sequence. First, Gaussian smoothing to attenuate high-frequency noise before differentiation amplifies it. Second, Sobel or similar gradient operators to compute magnitude and orientation at each pixel. Third, non-maximum suppression to thin multi-pixel-wide ridges down to single-pixel width by comparing each pixel to its neighbors along the gradient direction. Fourth, double thresholding to partition pixels into strong edges, weak edges, and non-edges using high and low thresholds. Fifth, edge tracking by hysteresis to retain weak pixels only if they are connected to strong pixels, which suppresses noise while preserving true contour continuity. The superiority claim should contrast this pipeline against naive Sobel thresholding: the latter yields thick edges because it lacks thinning, breaks contours because it lacks connectivity logic, and is less robust to noise because it lacks the hysteresis continuity test.

COMMON WRONG ANSWERS: Describing Canny as just Sobel plus Gaussian blur and a threshold. Omitting non-maximum suppression entirely or confusing it with simple thresholding. Treating hysteresis as a single threshold or as morphological dilation rather than connectivity-based tracking. Claiming Canny guarantees closed contours; it improves continuity but does not guarantee closure. Failing to mention that gradient orientation is required for non-maximum suppression, not just magnitude.

LIKELY FOLLOW-UPS: How do you choose the Gaussian sigma and the two thresholds in practice? What are the computational trade-offs versus modern learned edge detectors? How does the algorithm behave on low-contrast or texture-rich regions? Can you explain why non-maximum suppression uses interpolation between neighboring pixels rather than discrete directions? How would you adapt Canny to color or multispectral images?

ONE CONCRETE EXAMPLE: Consider an autonomous vehicle camera detecting lane markings. A naive Sobel threshold on the road image returns thick, blurred responses from asphalt texture and shadow boundaries. After Canny's Gaussian smoothing, texture noise is reduced. Sobel gradients capture the lane edges. Non-maximum suppression thins each lane boundary to a one-pixel-wide curve. Double thresholding discards weak shadow gradients while marking strong lane paint edges. Finally, hysteresis tracking connects short weak segments of a dashed lane line to the strong segments, producing a continuous navigable contour that a simple threshold would have fragmented.

Source: Wikipedia: Canny edge detector