Edges and regions turn raw pixels into structures that can support measurement, recognition, and segmentation.
Main idea
Edges describe where image values change. Regions describe which pixels belong together.
Complete Process
flowchart LR P["Pixels<br/>raw measurements"] --> E["Edge responses<br/>local change"] E --> L["Linked edges<br/>global curves"] L --> R["Regions<br/>pixels grouped together"] R --> S["Segmentation<br/>labels or masks"] P --> RF["Region features<br/>intensity, colour, texture"] RF --> R
Edge Branch
The edge branch starts with local intensity change and builds larger boundary structures:
- Estimate local change with Prewitt, Sobel, or another derivative filter
- Detect edgels by thresholding the response
- Thin and link edgels with methods such as Canny
- Find global shapes with methods such as the Hough Transform
Laplacian of Gaussian (LoG) uses smoothing and a second-order derivative to place an edge at a zero crossing.
Region Branch
The region branch groups pixels that share useful features:
- Intensity for brightness-based grouping
- Colour for separating surfaces with different colour values
- Texture for repeated local patterns
- Semantics for meaning-aware labels such as road, car, person, and sky
Otsu Thresholding separates two intensity classes and produces a binary mask. Morphological operations can then repair the mask through dilation, erosion, opening, or closing.
Which Structure Is Useful?
| Required result | Useful structure or method |
|---|---|
| Find rapid local intensity change | Image gradient |
| Obtain thin, connected boundaries | Canny edge detector |
| Convert many local edge pixels into a straight line | Hough transform |
| Describe a group by brightness or variation | Region mean and variance |
| Describe repeated local patterns | Texture and Gabor features |
| Split an image into two intensity classes automatically | Otsu thresholding |
| Assign a meaning to every pixel | Semantic segmentation |