CDS 6334 - Visual Image Processing

Lecture 7: Textures

1. What is Texture?

Texture refers to repeated local patterns in an image that describe surface appearance and material properties.
Textures can be regular (structured) or irregular (random).
🧠 Think:

Shape tells us what an object is.
Texture tells us what it is made of.

2. Why Analyze Texture?

Exam Keyword:
Texture Representation

3. Texture-Related Tasks

Task Description
Shape from Texture Estimate surface orientation or shape
Texture Segmentation Group similar texture regions
Texture Classification Recognize material types
Texture Synthesis Generate new texture images

4. Are Edges Enough?

Edge detectors capture boundaries but often fail to fully represent texture information.
Texture requires more information than simple edge locations.

5. Texture Representation

Texture can be represented using local statistics computed from image windows.
Common statistics:
  • Mean
  • Standard Deviation
  • Filter Responses
🧠 Texture = Pattern + Statistics

6. Window-Based Texture Analysis

Small neighbourhood windows are analyzed to summarize local texture patterns.
Each window becomes a feature vector describing its texture.

7. Texture Grouping

Similar texture windows can be grouped together into texture categories.
Example Texture Types:
  • Horizontal edges
  • Vertical edges
  • Low-gradient regions
  • Mixed patterns

8. Texture Distance

Texture similarity is measured using distances between feature vectors.
Important Formula:
Euclidean Distance
🧠 Smaller distance = More similar textures

9. Challenges in Texture Analysis

10. Filter Banks

A filter bank is a collection of filters used to capture different texture patterns.
Filters may detect edges, bars, spots, and structures at different orientations and scales.
🧠 One filter = Limited information
Filter Bank = Rich texture description

11. Leung-Malik (LM) Filter Bank

A popular texture filter bank consisting of 48 filters.
Includes:
  • 1st Derivative of Gaussian
  • 2nd Derivative of Gaussian
  • Laplacian of Gaussian (LoG)
  • Gaussian Filters
Exam Point:
48-dimensional feature vector

12. Filter Responses

Applying a filter bank produces multiple responses for each image location.
The collection of responses becomes a feature vector describing texture.

13. High-Dimensional Features

If a filter bank contains d filters, each texture feature vector has d dimensions.
Example:
48 filters → 48-dimensional feature vector

14. Clustering

Clustering groups similar feature vectors together.
Helps discover texture categories automatically.

15. K-Means Clustering

K-Means partitions data into K clusters by minimizing within-cluster variance.
Exam Keyword:
Sum of Squared Distances (SSD)

16. K-Means Algorithm

  1. Choose K
  2. Initialize cluster centres
  3. Assign points to nearest centre
  4. Compute cluster means
  5. Repeat until convergence
🧠 Assign → Update → Repeat

17. K-Means Properties

Advantages Disadvantages
Simple and fast Need to choose K
Guaranteed convergence Sensitive to initialization
Easy implementation Sensitive to outliers

18. Histograms of Texture Features

After clustering, texture occurrences can be summarized using histograms.
Histogram bins represent texture categories.

19. Textons

Textons are representative texture patterns discovered through clustering.
Process:
  1. Apply filter bank
  2. Cluster responses
  3. Create textons
  4. Build texton histograms
Important Term:
Universal Textons

20. Texture-Based Grouping

Pixels or regions with similar filter-bank responses are grouped together.
Useful for segmentation and classification.

21. Applications of Texture Analysis

22. Scene Classification

Images can be classified using texton histograms and similarity measures.
Similar scenes often have similar texture distributions.

23. Texture-Based Image Retrieval

Images are matched based on texture similarity rather than colour alone.
🧠 Colour may look similar,
Texture reveals the difference.

24. Chi-Square Distance

A common measure for comparing texture histograms.
Often performs better than Euclidean distance for histogram matching.

25. Final Exam Summary

Most Important Points

  • Texture: Repeated local image patterns.
  • Texture Analysis: Segmentation, classification, synthesis.
  • Texture Representation: Statistics from local windows.
  • Filter Bank: Multiple filters capture texture information.
  • LM Filter Bank: 48 filters.
  • Feature Vector: Collection of filter responses.
  • Texture Distance: Euclidean distance.
  • Clustering: Groups similar textures.
  • K-Means: Assign → Update → Repeat.
  • Textons: Representative texture patterns.
  • Histogram: Describes texture occurrence frequency.
  • Applications: Scene classification, texture matching, image retrieval.