CDS 6334 - Visual Image Processing

Lecture 10: Feature Indexing

1. Why Feature Indexing?

Large image databases contain millions of local features that must be searched efficiently.
Feature indexing speeds up image matching and retrieval.
🧠 Search smarter, not harder.

2. Matching Local Features

Candidate matches are generated by finding descriptors with similar appearance.
Brute-force matching compares every feature against all others.

3. Local Features in Feature Space

Each descriptor corresponds to a point in a high-dimensional feature space.
Example:
SIFT = 128-dimensional vector.

4. Similar Descriptors

Nearby points in feature space represent similar local image content.
🧠 Close points → Similar patches.

5. The Scalability Problem

Modern image collections may contain millions of descriptors.
Efficient indexing is required for large-scale retrieval.

6. Inspiration from Text Retrieval

Search engines use word indices to locate documents efficiently.
Similar ideas can be applied to image features.

7. Text Retrieval vs Image Search

Both problems can use a vocabulary and indexing strategy.
Exam Keyword:
Visual Vocabulary

8. Visual Words Concept

Image descriptors are converted into discrete visual words.
🧠 Features become words.

9. Visual Vocabulary

A collection of visual words forms a visual vocabulary.
Similar to a dictionary in text retrieval.

10. Clustering Descriptors

Similar descriptors are grouped together using clustering algorithms.
Common Method:
k-Means Clustering

11. Quantization

Quantization assigns each descriptor to its nearest cluster center.
🧠 Descriptor → Visual Word

12. Cluster Centers

Cluster centers serve as representatives of visual words.
Each center defines one word in the vocabulary.

13. Visual Word Assignment

Each image feature is assigned to the nearest visual word.
This converts continuous descriptors into discrete labels.

14. Textons

Textons are cluster centers derived from filter responses.
Used for texture and material representation.

15. Bag of Visual Words (BoVW)

An image is represented by the distribution of visual words.
Important Concept:
Bag of Visual Words

16. Histogram Representation

The occurrence count of each visual word is stored in a histogram.
🧠 Count words, not positions.

17. Analogy with Documents

Images are treated like documents and visual words like text words.

18. Comparing Images

Similar images have similar visual word distributions.
Compare histograms instead of individual descriptors.

19. Cosine Similarity

Measures similarity between two Bag-of-Words vectors.
Important Formula:
Cosine Similarity

20. Vocabulary Formation Issues

21. Inverted File Index

Maps visual words to images containing those words.
🧠 Word → Image List

22. Purpose of Inverted Index

Enables fast retrieval without scanning all images.
Widely used in large-scale image search systems.

23. Sparse Representation

Most visual words do not occur in a given image.
Sparse matrices save memory and computation.

24. tf-idf Weighting

Term Frequency – Inverse Document Frequency weighting improves retrieval performance.
Important Method:
tf-idf

25. Purpose of tf-idf

Frequent words within an image are emphasized while common words across many images are downweighted.
🧠 Rare words are more informative.

26. Limitation of Bag-of-Words

Spatial relationships between features are ignored.
BoW is an orderless representation.

27. Improving BoW

28. Application: Image Retrieval

Retrieve images or objects that match a query image.
Example:
Video Google System

29. Retrieval Evaluation

Precision and Recall are used to evaluate retrieval quality.
Exam Favourite:
Precision = Relevant Retrieved / Retrieved
Recall = Relevant Retrieved / Total Relevant

30. Final Exam Summary

Most Important Points

  • Visual Words: Quantized local descriptors.
  • k-Means: Creates visual vocabulary.
  • Quantization: Assign descriptors to nearest cluster.
  • Bag of Visual Words: Histogram of word occurrences.
  • Cosine Similarity: Compare image histograms.
  • Inverted File Index: Fast image retrieval.
  • tf-idf: Weight important visual words.
  • BoW Limitation: Ignores spatial relationships.
  • Applications: Image retrieval and object search.
  • Evaluation: Precision and Recall.