2. Contents
Introduction to image retrieval
Introduction to texture
Methods of Extracting
Evaluation of some approaches and results
References
2
3. Content-based Image Retrieval
An image search engine Works by image content (Color, texture,
shape) instead of annotated texts
Consist of:
1. Database of Primitive Image
2. Feature Extraction Module
3. Indexing Module
4. Search and Retrieval Module
5. User interface (Input: query image, Output: Similar images)
Examples: IBM QBIC, Virage, VisualSEEK, …
3
4. CBIR
CBIR modules and flowchart
Row Images
Query Image
Feature
Extraction
Feature
Vectors
Feature
Vector DB
Feature
Vector
Feature
Vector
Search and
retrieval
Similarity
Measure
Results to
User
4
5. What is texture?
A key component of human visual perception about the nature
and three dimensional shape of physical objects
Can be regarded as a similarity grouping in an image
One of essential Features to consider when querying image
database
Normally defined by grey levels
5
6. Texture analyzing
Rottenly :
it is required to convert image into gray scald mode
Inspecting batch of pixels in order to find the relationship
between them
6
7. Methods of analyzing
Approaches to texture analysis are usually categorized into
Structural,
Statistical,
Model-based and
Transform
7
8. Structural approaches
Represent texture by well-defined primitives called
microtexture and a hierarchy of spatial arrangements of those
primitives
Define the primitives and the placement rules to define the
texture
8
9. Statistical approaches
Represent the texture
indirectly by the non-
deterministic properties
These properties govern the
distributions and
relationships between the
grey levels of an image
Approaches
Edge Histogram Descriptor
Co-occurrence matrix
Tamura features
9
10. Model-based approaches
Attempt to interpret an
image texture by use of,
respectively, generative
image model
Approaches
Auto-regressive (AR)
Gaussian-Markov RMF
Gibbs RMF
10
11. Transform approaches
Represent an image in a
space whose co-ordinate
system (such as frequency
or size)
Interpretation in this space
will be closely related to the
characteristics of its texture
Approaches
Gabor Transform
Wavelet Transform
11
12. Problem & Experimental Set up
To Evaluate three texture extraction method to use
in Content-based Image Retrieval
Image Collection: Corel Collection
Similarity Measure: Manhattan Metric
12
13. Co-occurrence matrix
Definition
One of the earliest methods
Also called GLCM stands for Gray-level Co-occurrence Matrix
Extract 2nd-order statistics from an image
Very successful method
13
14. Co-occurrence matrix (cont.)
Let C be the GLCM, so Ca,d(i,j) will be the co-occurrence of
pixels with grey values i and j at a given distance d and in a
given direction α
Should be symmetric or asymmetric
Usually:
All pixel intensities are quantized into smaller number of available gray
levels (8, 16, 64, …). For example if 8 is selected, the target matrix will be
8 x 8.
Values of α are one of values such as 0, 45, 90 and 135. Using all of them
may bring more accuracy.
14
16. Co-occurrence matrix (cont.)
Feature Extraction:
Once the GLCM has been
created, various features
can be computed from
it.
All these features are
supported by MATLAB
Formula
Feature
Energy
Entropy
Contrast
Homogeneity
16
17. Co-occurrence matrix – Evaluation Results
Distance between 1 and 4 pixels gave the best performance
There was no significant differences between symmetrical and
asymmetric matrices
Tiling of the image gave a large increase in retrieval which
flatted out by 9 x 9 tiles
The concatenated (cat) features gave better result at all points
than the rotationally invariant summed matrices (sum)
The best feature was homogeneity
17
19. Tamura
Extract features that correspond to human perception
Contains six textural features:
1. Coarseness
2. Contrast
3. Directionality
4. Line-likeness
5. Regularity
6. Roughness
19
20. Tamura (cont.)
20
First three are most important
Coarseness
direct relationship to scale and repetition rates
calculated for each points of image
Contrast
dynamic range of gray levels in an image
calculated for non-overlapping neighborhoods
Directionality
Measure the total degree of directionality
calculated for non-overlapping neighborhoods
21. Tamura (cont.)
21
Another approach: Tamura CND Image
Spatial joint of coarseness-contrast-directionality distribution (view as
RGB distribution)
Extract color histogram style feature from Tamura CND Image
22. Tamura – Evaluation Results
22
Increasing k value for coarseness decrees the performance
Optimum value = 2
Performance of directionality is poor
24. Gabor filter
Special case of the short-time Fourier transform
Time-frequency analysis
It is used to model the responses of human visual system
A two dimensional Gabor function
Advantage/disadvantage:
Very popular
Time consuming calculation
Generate complete but non orthogonal basic set so redundancy of data
will be occurred
24
25. Gabor filter (cont.)
Manjunath et al reduced redundancy by using Gabor wavelet
functions
The Features is computed by
1. Filtering the image with a bank of orientation and scale sensitive
filters and,
2. Computing the mean and standard deviation of the output in the
frequency domain
25
26. Gabor filter – Evaluation Results
26
Better for homogeneous textures with fixed size because of specific filter
dictionary
Widely used to search for an individual texture tile in an aerial images
database
Best response Usage:
Process image for 7x7 tiling and apply filters on
Just 2 scales and 4 orientations
28. References
1) Howarth P. and Ruger S., "Evaluation of Texture Features for Content-
Based Image Retrieval," in Third International Conference, CIVR 2004,
Dublin, Ireland, 2004.
2) Deselaers Th., "Features for Image Retrieval," 2003
3) Materka A. and Strzelecki M. , "Texture Analysis Methods – A Review,"
Technical University of Lodz, Institute of Electronics, Brussels, COST B11
1998.
4) Manjunath B.S. and Ma W.Y., "Texture features for browsing and retrieval
of image data," Transactions on Pattern Analysis and Machine
Intelligence, vol. 18, pp. 837-842, 1996.
5) Schettini R. ; Ciocca G. and Zuffi S., "A Survey of Methods for Color Image
Indexing and Retrieval in Image Databases".
28
29. Appendix: Performance measures of an Information
Retrieval System
Every document is known to be either relevant or non-relevant to a particular query
1. Precision: The fraction of the documents retrieved that are relevant to the user's
information need
Precision = (Relevant images ∩ Retrieved Images) / Retrieved Images
2. Recall: The fraction of the documents that are relevant to the query that are
successfully retrieved
Recall = (Relevant images ∩ Retrieved Images) / Relevant images
3. Average Precision: The precision and recall are based on the whole list of documents
returned by the system. Average precision emphasizes returning more relevant
documents earlier. It is average of precisions computed after truncating the list after
each of the relevant documents in turn:
AveP = Ʃr = 1:n (P(r) . rel(r)) / Relevant images
where r is the rank, N the number retrieved, rel() a binary function on the relevance of a
given rank, and P() precision at a given cut-off rank.
29