Canny isn't one kernel — it's four stages working in sequence, each fixing a weakness of the one before it. Blur tames noise. Gradient finds where brightness changes. Non-max suppression thins the fat gradient response down to single-pixel lines. Hysteresis decides which of those lines are real. Walk through every stage below, then watch them run back to back.
Every edge detector after this one operates on this blurred image, not the original. Too little σ and noise survives to corrupt the gradient; too much and real fine edges get washed away before they're even measured. Canny's default is a mild σ≈1.4 — enough to matter, not enough to erase detail.
Gradients (Stage 2) are derivatives, and derivatives amplify noise — exactly the problem Lab 11 showed with raw Laplacian. Blurring before differentiating (this is literally the LoG idea again) keeps the gradient stage from drowning in false responses.
Sobel turns "brightness" into "strength + direction" at every pixel. Without direction, Stage 3 would have nothing to suppress along — magnitude alone can't tell which neighbours are "across" the edge versus "along" it.
Raw gradient magnitude is a blurry ridge several pixels wide around every real edge. Non-max suppression walks along the gradient direction at each pixel and keeps only the single local peak — collapsing that ridge to a one-pixel-wide line.
A single fixed threshold either lets noise through (too low) or breaks real edges into dashes (too high). Two thresholds plus connectivity gets the best of both — confident edges anchor the decision, borderline pixels are only kept if they're actually attached to something real.
You set the high threshold very low (barely above the low threshold) and hit Trace Connectivity on a photo with visible sensor noise. What's the most likely outcome?