Image segmentation is the process of partitioning a digital image into multiple meaningful regions or objects. These regions represent sets of pixels with similar attributes like intensity, color, or texture.
Let the entire spatial region of the image be denoted as R. The goal of segmentation is to divide R into n subregions: R1,R2,R3,…,Rn
These subregions must satisfy the following five conditions:
1. Completeness:
- Every pixel in the image must be assigned to a region.
- No pixel should be left unclassified.
2. Connectivity:
- Each region Ri must be a connected set of pixels.
- Pixels in a region must be linked based on some connectivity rule (e.g., 4-connected or 8-connected).
3. Disjointness:
- No two regions should overlap.
- Each pixel belongs to only one region.
4. Homogeneity Predicate:
- Each region must satisfy a property Q.
- Example: All pixels in RiR_i have the same or similar intensity, texture, or color.
5. Maximality Between Regions:
- No two adjacent regions should satisfy predicate Q when combined.
- This ensures regions are maximally distinct according to Q.
Approaches to Image Segmentation
Image segmentation techniques are generally classified into two categories:
1. Discontinuity-based Segmentation (Edge-based):
- Based on sudden changes in intensity (edges).
- Assumes that region boundaries produce strong intensity contrasts.
- Common methods: Edge detection, Gradient-based methods
2. Similarity-based Segmentation (Region-based):
- Groups together pixels that are similar in intensity, color, or texture.
- Based on region growing, thresholding, clustering, etc.
Example with Figure 10.1 (Concept Explanation):
(a) & (b) Edge-based Segmentation:
- An image with a constant region on a darker background.
- Edge detection finds the boundary between object and background.
- All pixels inside boundary are labeled one value (e.g., white), and outside another (e.g., black).
(c) Region-based Segmentation:
- Assigns uniform intensity to each segmented region.
- Satisfies all five segmentation conditions.
(d)–(f) Textured Image Segmentation:
- (d) Image with textured inner region.
- (e) Edge detection fails due to many false edges.
- (f) Region-based method using standard deviation:
- Divide the image into 8×8 blocks.
- If the standard deviation > 0, label as textured (white), else constant (black).