Morphological Image Processing: Erosion, Dilation, Opening, Closing & Applications

TAKEAWAYS:

• Morphological operations use structuring elements to probe and modify images via set theory (union, intersection, complement).
• Erosion shrinks objects (removes borders) while dilation expands them (fills gaps).
• Opening (erosion + dilation) removes small noise; closing (dilation + erosion) fills small holes.
• Hit-or-miss transforms detect specific patterns (e.g., corners, lines) by combining erosion with two structuring elements.
• Applications include license plate recognition (erosion to separate characters), medical imaging (segmenting tumors), and satellite cloud removal.
• Trade-offs: small structuring elements preserve details but may under-segment; large ones smooth but lose fine structures.

Core Concepts: Morphology as Set Theory on Images

Morphological image processing treats an image as a set of pixels (e.g., foreground = 1, background = 0) and applies set operations using a structuring element (SE). The SE acts like a "probe" or "stamp" that slides over the image, modifying it based on overlaps.

1. Structuring Elements (SEs)

The SE defines the neighborhood for operations. Common shapes:

• Flat SEs: Binary (0/1) or grayscale (intensity values).
• Shapes: Disk, square, cross, line segment, or custom (e.g., a "T" shape for detecting junctions).

Example SEs:

graph LR
    SE1["Disk-shaped SE"] -->|"Shape"| circle("30")
    SE2["Square SE"] -->|"Shape"| square("30")
    SE3["Cross-shaped SE"] -->|"Shape"| plus("30")

Real-world SE choice:

• Medical imaging: Disk-shaped SEs for tumor segmentation (smooth boundaries).
• OCR (e.g., eSewa forms): Cross-shaped SEs to separate handwritten digits.

2. Basic Operations: Erosion and Dilation

Erosion (θ)

• Definition: A pixel in the output is 1 only if all pixels in the SE are 1 in the input.
• Effect: Shrinks foreground objects, removes small noise, and separates touching objects.
• Mathematically: $$(A\ominus B)(x,y)=\bigcap _{(x^{\prime },y^{\prime })\in B}A(x+x^{\prime },y+y^{\prime })$$ where $A$ = image, $B$ = SE.

Worked Example: Erosion with a 3×3 Square SE Input Image (3×3):

1 1 1
1 1 0
1 0 0

SE (3×3 square):

1 1 1
1 1 1
1 1 1

Output:

• Center pixel (1,1): SE overlaps (0,0) to (2,2). All 9 pixels in input? No (input has 0 at (2,1) and (2,2)).
• Result: Only the top-left 1 survives (since its SE fits entirely in the input’s 1s).

1 0 0
0 0 0
0 0 0

Visual Trace:

Dilation (⊕)

• Definition: A pixel in the output is 1 if it overlaps with any 1 in the SE.
• Effect: Expands foreground objects, fills gaps, and connects nearby objects.
• Mathematically: $$(A\oplus B)(x,y)=\bigcup _{(x^{\prime },y^{\prime })\in B}A(x-x^{\prime },y-y^{\prime })$$

Worked Example: Dilation with a 3×3 Square SE Input Image (same as above):

1 1 1
1 1 0
1 0 0

Output:

• Any pixel whose SE overlaps at least one 1 becomes 1.
• Result:

1 1 1
1 1 1
1 1 1

Comparison Table:

3. Advanced Operations: Opening and Closing

Opening (∘)

• Definition: Erosion followed by dilation (A ∘ B = (A ⊖ B) ⊕ B).
• Effect:
  • Removes small noise and fine details.
  • Preserves the shape and size of larger objects.
• Example: Cleaning up text in a scanned document (e.g., Daraz’s product images).

Worked Example: Opening on Noisy Text Input (noisy "A"):

0 1 0
1 1 1
0 1 0

After Erosion (3×3 SE):

0 0 0
0 1 0
0 0 0

After Dilation:

0 1 0
1 1 1
0 1 0

Result: Small noise removed; "A" preserved.

Closing (•)

• Definition: Dilation followed by erosion (A • B = (A ⊕ B) ⊖ B).
• Effect:
  • Fills small holes and gaps.
  • Connects nearby objects.
• Example: Restoring broken lines in satellite images (e.g., NTC’s road network maps).

Worked Example: Closing a Broken Line Input (broken line):

1 1 1
1 0 1
1 1 1

After Dilation (3×3 SE):

1 1 1
1 1 1
1 1 1

After Erosion:

1 1 1
1 1 1
1 1 1

Result: Gap filled.

Mermaid Diagram of Operations:

flowchart LR
    A["Original Image"] --> B["Erosion"]
    B --> C["Opening\n(A ⊖ B ⊕ B)"]
    A --> D["Dilation"]
    D --> E["Closing\n(A ⊕ B ⊖ B)"]

4. Hit-or-Miss Transform: Detecting Specific Patterns

• Definition: Detects pixels that match both a foreground SE (B) and a background SE (C). $$(A\otimes (B,C))(x,y)=(A\ominus B)(x,y)\cap (A^c\ominus C)(x,y)$$ where $A^c$ = complement of $A$.

• Use Case: Finding corners, line endings, or specific shapes (e.g., "T" junctions in Kathmandu traffic maps).

Example: Detecting Corners

• Foreground SE (B):

  0 1 0
  1 1 1
  0 1 0
  

• Background SE (C):

  1 1 1
  1 0 1
  1 1 1
  

• Result: Pixels where the foreground matches B and the background matches C are marked as corners.

Real-world Application:

• Pathao’s route optimization: Detects "dead ends" (corners) in Kathmandu’s narrow streets using hit-or-miss transforms to avoid traffic jams.

5. Grayscale Morphology

For grayscale images, operations compare pixel values to the minimum/maximum in the SE neighborhood:

• Erosion: Replace pixel with the minimum value in the SE.
• Dilation: Replace pixel with the maximum value in the SE.

Example: Smoothing a Grayscale Image Input (3×3 grayscale patch):

50 60 70
40 50 60
30 40 50

Erosion (3×3 SE):

• Center pixel (1,1) = min(50,60,70,40,50,60,30,40,50) = 30. Dilation:
• Center pixel = max(...) = 70.

Effect: Erosion darkens edges; dilation brightens them.

## In the Real World

1. eSewa’s License Plate Recognition

   • Operation: Opening (erosion + dilation) with a disk-shaped SE to separate overlapping characters on Nepal’s license plates.
   • Why? Erosion breaks apart touching digits (e.g., "8" and "B"), while dilation reconnects the skeleton of each character.

2. NTC’s Satellite Imagery for Flood Mapping

   • Operation: Closing with a large SE to fill small gaps in cloud-covered areas, revealing submerged roads.
   • Example: During monsoon floods, closing helps distinguish rivers from clouds in raw satellite data.

3. Khalti’s Document Verification

   • Operation: Hit-or-Miss Transform to detect forged signatures by matching known patterns (e.g., "Khalti" logo corners).
   • How? The transform flags pixels where the foreground matches the logo’s shape and the background matches the expected empty space around it.

4. Medical Imaging: Tumor Segmentation

   • Operation: Morphological Gradient (A ⊕ B - A ⊖ B) to highlight edges of brain tumors in MRI scans.
   • Example: At Tribhuvan University Hospital, erosion with a disk SE shrinks the tumor region, while dilation expands it; subtracting the two enhances the boundary.

## Worked Example: Traffic Sign Segmentation

Scenario: Segmenting circular traffic signs (e.g., "No Entry") from a blurry road image. Steps:

1. Convert to binary: Threshold at gray level = 128.
2. Erode with disk SE (radius=5): Removes small noise (e.g., dust on the lens).
3. Dilate with same SE: Reconnects broken sign edges.
4. Hit-or-Miss with circular SE: Detects only circular objects (ignores rectangular signs).

Input Image (simplified):

1 1 1 1 1
1 0 0 0 1
1 0 1 0 1
1 0 0 0 1
1 1 1 1 1

After Erosion (3×3 disk SE):

0 0 0 0 0
0 0 0 0 0
0 0 1 0 0
0 0 0 0 0
0 0 0 0 0

After Dilation:

0 0 1 0 0
0 1 1 1 0
0 1 1 1 0
0 0 1 0 0
0 0 0 0 0

Result: Only the circular sign’s skeleton remains.

## Exam Tip

1. Memorize Definitions:

   • Erosion = "shrinks," Dilation = "expands."
   • Opening = "remove small noise," Closing = "fill small holes."
   • Hit-or-Miss = "detect specific patterns."

2. Structuring Element Choice:

   • Square/disk SEs: For isotropic smoothing (e.g., medical images).
   • Cross-shaped SEs: For directional operations (e.g., separating text lines).
   • Always state the SE size/shape in answers (e.g., "3×3 square SE").

3. Common Pitfalls:

   • Order matters: Opening ≠ Closing. Always erosion first, then dilation.
   • Grayscale vs. Binary: Grayscale uses min/max; binary uses set operations.
   • Edge artifacts: Erosion can remove entire small objects; dilation can merge nearby objects.

4. Exam Questions Pattern:

   • Theory (30%): Define erosion/dilation, explain opening/closing.
   • Calculation (40%): Given an image and SE, compute erosion/dilation manually (use 3×3 examples).
   • Application (30%): Match operations to real-world tasks (e.g., "How would you clean noise from a Daraz product image?" → Opening).

5. Quick Trick for Manual Calculation:

   • For erosion: All pixels in the SE must be 1 in the input.
   • For dilation: At least one pixel in the SE must be 1 in the input.
   • Use a template overlay method (draw the SE over the image and check overlaps).