Morphology Toolkit

When to Use This Skill

• Cleaning binary images after thresholding (removing noise, filling holes)
• Separating touching objects in segmentation results
• Extracting object boundaries or skeletons
• Enhancing contrast for specific features (Top-Hat/Bottom-Hat)
• Processing fingerprints, cell images, document scans

Decision Framework

The Core Four Operations

What do you need to clean up?
├── Small white noise specks on black background (external noise)
│   └── ✅ Opening (Erosion → Dilation)
│
├── Small black holes inside white objects (internal gaps)
│   └── ✅ Closing (Dilation → Erosion)
│
├── Objects touching each other (need separation)
│   └── ✅ Erosion (shrinks objects apart)
│
└── Broken lines or gaps in object boundaries
    └── ✅ Dilation (expands and connects)

Opening vs Closing — The Memory Trick

| Operation | Formula | Removes | Keeps | Memory Aid | |-----------|---------|---------|-------|------------| | Opening | Erode → Dilate | External noise | Object integrity | "Opens" gaps between noise and object | | Closing | Dilate → Erode | Internal holes | Object shape | "Closes" holes inside object |

> Golden rule: Opening cleans the outside, Closing fills the inside.

Real-world example — Fingerprint processing:

1. Opening removes background dirt/smudges
2. Closing reconnects broken ridge lines

Structuring Element Selection

| Shape | OpenCV Constant | Best For | |-------|----------------|----------| | Rectangle/Square | cv2.MORPH_RECT | Angular objects, text characters | | Ellipse/Disk | cv2.MORPH_ELLIPSE | Round objects, cells, coins | | Cross (+) | cv2.MORPH_CROSS | Thin lines, intersections |

Size selection rule: The structuring element must be:

• Larger than the noise you want to remove
• Smaller than the objects you want to keep

python
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))

Contrast Enhancement Morphology

| Filter | Formula | Reveals | Use Case | |--------|---------|---------|----------| | Top-Hat | Original - Opening | Bright details on dark background | Bright spots, text on dark surface | | Bottom-Hat | Closing - Original | Dark details on bright background | Dark spots, stains on light surface |

python
tophat = cv2.morphologyEx(gray, cv2.MORPH_TOPHAT, kernel)
blackhat = cv2.morphologyEx(gray, cv2.MORPH_BLACKHAT, kernel)

Critical Gotchas

1. Order Matters (Opening ≠ Closing)

Swapping the order of erosion and dilation produces opposite results:

• Erode first, then Dilate = Opening (removes small white blobs)
• Dilate first, then Erode = Closing (fills small black holes)

2. Grayscale Morphology Works Differently

In binary images, erosion/dilation work with set theory. In grayscale:

| Operation | Grayscale Behavior | Visual Effect | |-----------|-------------------|---------------| | Dilation | Maximum filter (picks brightest neighbor) | Image brightens, bright regions expand | | Erosion | Minimum filter (picks darkest neighbor) | Image darkens, dark regions expand |

3. Iteration Count Amplifies Effect

python
# Single pass
eroded = cv2.erode(binary, kernel, iterations=1)

# Multiple passes = stronger effect (equivalent to larger kernel)
eroded = cv2.erode(binary, kernel, iterations=3)

4. Morphological Gradient = Edge Extraction

python
# Dilation - Erosion = object boundary
gradient = cv2.morphologyEx(binary, cv2.MORPH_GRADIENT, kernel)

Quick Reference

Basic Operations

python
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))

eroded = cv2.erode(binary, kernel, iterations=1)
dilated = cv2.dilate(binary, kernel, iterations=1)
opened = cv2.morphologyEx(binary, cv2.MORPH_OPEN, kernel)
closed = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel)

Advanced Operations

python
gradient = cv2.morphologyEx(img, cv2.MORPH_GRADIENT, kernel)  # Boundary
tophat = cv2.morphologyEx(img, cv2.MORPH_TOPHAT, kernel)      # Bright details
blackhat = cv2.morphologyEx(img, cv2.MORPH_BLACKHAT, kernel)   # Dark details

Common Processing Chain

python
# Typical binary cleanup pipeline
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
cleaned = cv2.morphologyEx(binary, cv2.MORPH_OPEN, kernel)   # Remove noise
cleaned = cv2.morphologyEx(cleaned, cv2.MORPH_CLOSE, kernel) # Fill holes