Edge Detection

Image Processing with Edge Detection

Edge Detection is a fundamental image processing and classical computer vision technique that aims to identify points in a digital image where luminous intensity (brightness) changes abruptly. These changes generally correspond to physical object boundaries.

⚙️ How it works: Practical Context

Mathematically, edges are detected by calculating the image’s gradient, which measures the direction and rate of change in pixel intensity.

Most common classical algorithms:

• Sobel: Applies simple convolution matrices to find vertical and horizontal gradients. Fast, but produces thick edges.
• Canny Edge Detector: A more complex pipeline involving:
  1. Gaussian Blur to remove noise.
  2. Sobel gradient calculation.
  3. Non-Maximum Suppression to keep only the local maximum “ridge” of the edge (thinning it).
  4. Hysteresis Thresholding to discard false edges and connect true ones.

🧮 Mathematical Foundation (Image Gradient)

Given an image represented by a matrix $A$, the Sobel operator approximates the derivatives using two $3 \times 3$ convolution kernels, one detecting horizontal changes ($G_x$) and the other vertical changes ($G_y$):

$$G_x = \begin{bmatrix} -1 & 0 & +1 \\ -2 & 0 & +2 \\ -1 & 0 & +1 \end{bmatrix} * A \quad \text{and} \quad G_y = \begin{bmatrix} -1 & -2 & -1 \\ 0 & 0 & 0 \\ +1 & +2 & +1 \end{bmatrix} * A$$

Eq. 1: Sobel operators for horizontal and vertical gradients.

The matrix $G_x$ is designed to respond maximally to vertical edges, and $G_y$ to horizontal edges. The symbol $*$ denotes the 2D discrete convolution operation.

At each pixel, we calculate the gradient magnitude (the edge’s “strength”) by combining these two approximations:

$$|G| = \sqrt{G_x^2 + G_y^2}$$

Eq. 2: Gradient magnitude.

Where:

• $|G|$: total gradient magnitude at the pixel.
• $G_x$: gradient component in the horizontal direction.
• $G_y$: gradient component in the vertical direction.

And the edge orientation (angle), which is crucial for the Non-Maximum Suppression step in Canny:

$$\theta = \arctan\left(\frac{G_y}{G_x}\right)$$

Eq. 3: Edge direction (angle).

Where:

• $\theta$: edge orientation angle relative to the horizontal axis.
• $\arctan$: arc tangent, used to compute the angle from the gradient vector components.

With $\theta$, the algorithm checks the neighborhood along the gradient direction and discards the current pixel if it isn’t a local maximum, producing perfectly delineated thin edges.

🎯 When to use

• Preprocessing for ML: Reduces data volume. Instead of a heavy RGB image, an object-detection model might receive the edge map to focus on geometry.
• OCR (Optical Character Recognition).
• Classical feature extraction for systems without Deep Learning.

⚠️ Trade-offs and Pitfalls: Because the gradient measures sudden variations, it is highly sensitive to high-frequency noise. Applying derivatives to un-smoothed images will result in false edges scattered everywhere. A Gaussian Blur is always applied beforehand.

Related: object-detection