Q22Digital Image Processing

An Edge in an image is an absolute physical boundary indicating a catastrophic, instantaneous jump in pixel intensity (e.g., a black car against a white sky). Edge Detection algorithms mathematically identify these boundaries by executing 2D spatial derivatives. Because an image is a discrete matrix, algorithms use Finite Differences (Gradient Masks like Sobel or Prewitt) to aggressively calculate the Magnitude (edge strength) and Direction of the gradient at every single pixel.

Primitive detectors like Sobel are horrific; they produce extremely thick, blurry edges and are catastrophically vulnerable to thermal noise. In 1986, John Canny engineered the absolute optimal mathematical edge detector. It is not a simple mask; it is a massive, highly aggressive multi-stage algorithmic pipeline designed to minimize error rate, perfectly localize the edge, and guarantee a strict 1-pixel-thick response.

The 4-Stage Execution Pipeline

• 1. Gaussian Smoothing (Noise Annihilation): Thermal noise generates massive false gradients. The Canny algorithm first violently blurs the image using a Gaussian Low-Pass filter. This mathematically annihilates the high-frequency noise while preserving the massive structural edges.
• 2. Gradient Calculation (Sobel): The algorithm applies vertical and horizontal derivative masks (usually Sobel) to the smoothed image to calculate the absolute Gradient Magnitude ($M$) and Gradient Angle ($\theta$) for every single pixel.
• 3. Non-Maximum Suppression (Edge Thinning): The resulting edges are still heavily blurred and thick. Canny forces a strict 1-pixel thickness. The algorithm analyzes the gradient angle $\theta$ at a pixel, and aggressively looks at the two immediate neighbors strictly along that direction. If the center pixel's magnitude is NOT the absolute local maximum compared to its neighbors, its value is violently crushed to zero (black). This mathematically thins the edge to a perfect single-pixel line.
• 4. Hysteresis Thresholding (Edge Linking): Standard single thresholding causes edges to shatter into broken dashed lines. Canny utilizes a brilliant dual-threshold architecture (a High Threshold $T_H$ and a Low Threshold $T_L$).
• - If a pixel $>T_H$, it is mathematically locked as a "Strong Edge".
• - If a pixel $<T_L$, it is violently rejected.
• - If a pixel is in the middle, it is a "Weak Edge". The algorithm recursively traces the weak edge; if it physically connects to a Strong Edge, it is aggressively upgraded and saved. If not, it is annihilated. This perfectly preserves continuous boundaries while completely eradicating isolated noise.

Comparison with Sobel/Prewitt

A raw Sobel or Prewitt operator applied without pre-smoothing produces edges that are several pixels thick along a gradual intensity ramp, and any sensor noise generates scattered false-positive edge pixels throughout flat regions, since the operator responds to any local gradient above zero rather than the true underlying structural boundary. Canny's Gaussian smoothing stage suppresses this noise before the gradient is even computed, and non-maximum suppression collapses the resulting thick gradient ridge down to a single pixel by keeping only local maxima along the gradient direction. Hysteresis thresholding then solves the fragmentation problem that plagues single-threshold approaches: a lone global threshold either lets noise through (if set too low) or breaks genuine edges into disconnected dashes wherever contrast dips slightly (if set too high), whereas Canny's dual thresholds keep strong edges unconditionally and rescue weak edges only when they connect to a strong one, giving continuous boundaries without noise leakage. This is why Canny is regarded as the theoretically optimal edge detector under the criteria of good detection, good localization, and single response to a single edge, and remains the default choice in applications like the Hough-transform-based line detection used for boundary detection in later processing stages.