Evaluation of Content-Based Image Retrieval Systems
Based on Feature Extraction
Presented by
Fouzia Ashraf Khan
Motivation
 Exponential increase in computing power and electronic storage capacity
 Exponential increase in digital image/video database sizes
 Increase use of image and video:
Entertainment
Education
Commercial purposes
 Need abstractions for efficient and effective modeling
 Continuous data, multiple features and analysis of multi-layer
or multivariate data processing is difficult job.
Spatial Database
 Spatial Data
 Displayed, manipulated and analyzed by means of spatial attribute
which denotes a location on or near the surface of earth
Spatial attribute is presented in the form of co-ordinate pairs that
allow position and shape of a particular feature to be measured and
represented graphically.
• Spatial Database
A spatial database is a database that is optimized to store and query
data related to objects in space, including points, lines and polygons
Typical databases understand various numeric and character data
Need additional functionality spatial data types - called geometry or
feature.
CBIR
• Content-Based Image Retrieval
- a technique that utilizes the visual content of an image, to
search for similar images in large-scale image databases,
according to a user’s interest
- is a robust system based upon a combination of higher-
level and lower-level vision principles.
- Higher-level analysis uses perceptual organization,
inference and grouping principles to extract semantic
information
- Lower-level analysis employs a channel energy model to
describe image texture, shape and utilizes color histogram
techniques
CBIR System Components
 Three Major Components
 Feature Extraction
 Indexing
 System Design
Challenge:
Gap between low-level features and high level user
semantics
Feature Extraction
 Feature
A cluster of points or a boundary /region of pixels
 Primary Features
◦ Color
◦ Texture
◦ Shape
◦ Spatial location
• Feature extraction
Process of studying and locating areas and objects on the ground and deriving
useful information from images.
 Feature Extraction Methods
◦ Fuzzy approach
◦ Relevance feedback (supervised learning)
Example of a Classified Image
Color Based Feature Extraction
Approaches
 Conventional Color Histogram (CCH)
 Frequency of occurrence of every color in image.
 Refers probability mass function
 Captures the joint probabilities of intensities of color channels.
 hA,B,C(a,b,c) = N. Prob(A=a, B=b, C=c), where A, B and C are the three color channels and
N is the number of pixels in the image
 Fuzzy Color Histogram (FCH)
 Considers degree of color similarity between pixels
 A color space with K color bins, FCH of I is defined as
F(I)=[f1,f2,…fk] where fi = and
where N = # of pixels
μij = Membership value of jth pixel to ith color bin
dij = Euclidean distance between the color of pixel jth and the ith color bin,
ς = Average distance between colors in quantized color space

N
i
ij
N
)(
1


 dijxfij


1
1
)(
Color Based Feature Extraction
Approaches
 Color Correlogram (CC)
 Express how spatial correlation of pairs of colors changes with distance.
 A table indexed by color pairs
 For dth entry at location (i,j) is computed by
- Counting number of pixels of color j at a distance d from a pixel of color i,
divided by the total number of pixels in the image
 Color-Shape Based Method
 Includes area and shape information
 Area of each object is encoded as the number of pixels.
 Shape is characterized by “perimeter intercepted lengths”
 PIL - obtained by intercepting object perimeter with 8 line segments having 8
different orientations and passing through object center.
Comparative Study of Color Based
Feature Extraction
Texture Based Feature Extraction
Approaches
 Steerable Pyramid
◦ Basic filters are translation and rotation of a single function
◦ Filter is linear combination of basis functions
◦ Only for rotation-invariant texture retrieval
 Contourlet Transform
◦ Combination of a Laplacian pyramid and directional filter bank
◦ LP provides the multi‐scale decomposition,
◦ DFB provides the multi‐directional decomposition
 Gabor Wavelet
◦ Optimally achieves joint resolution in space and spatial frequency
◦ Transform dilates and rotates the two‐dimensional Gabor function.
◦ Image is convolved with each of the obtained Gabor functions
◦ Gabor function works in Fourier domain,
◦ Computationally intensive
 Complex Directional Filter Bank (CDFB)
◦ Retrieval results comparable with Gabor wavelet results
◦ Shift Invariant
Comparative Study of Texture
Based Feature Extraction
Shape Based Feature
Extraction Approaches
 Blob Detection
 Detecting points , regions are either brighter or darker than surrounding
 Two classes of blob detectors
- Differential method based on derivative expressions
- Methods based on local extrema of intensity
• Thresholding
 Classify pixels into two categories:
– Those to which some property measured from the image falls below a threshold,
and those at which the property equals or exceeds a threshold.
– Thresholding creates a binary image : binarization
- Choosing a threshold is a critical task.
 Hough Transform
- Find imperfect instances of objects within a certain class
of shapes by a voting procedure
- Voting procedure is carried out in a parameter space
- Object candidates are obtained as local maxima
 Template Matching
-For finding small parts of an image which match a template
-To detect edges in images.
-Simple to implement and understand, it is one of the slowest
methods.
High Dimensional Indexing
Techniques
 To make CBIR truly scalable to large size image collections
 Some Multi-dimensional Indexing Methods
 Self-Organizing Maps
- SOM are feed-forward, competitive artificial neural networks.
- Reduces complexity (both time and computation)
 Shape-Depth Representation
-Well-defined geometric objects, a novel strategy based on a 3D embedding
-Is between contour-based and region-based representations
-A very useful one for retrieving objects/images
 Fourier Transform of Shape-Depth and Indexing
- Invariant to transformations such as translation, scaling and rotation.
- Fourier transform is widely used for achieving the invariance
System Design
 Some factors of Design
- Random browsing
- Search by example
- Search by sketch
- Search by text (including keyword or speech) and navigation with
customized image categories.
Some Current Developed CBIR Systems
QBIC - Query By Image Content
Virage
RetrievalWare
VisualSEEk and WebSEEk
Netra
Performance Evaluation of CBIR System
Performance
 Observed Probability
 Distance Measure and Retrieval Score
 measure
- user is facing a target search task from a database D for an image I
belonging to class C D.
- Resembles the “target testing" method
- Obtained by implementing an “ideal screener“
- A computer program which simulates the human user by examining the
output of the retrieval system
- Marking the images returned by the system either as relevant (positive) or
non-relevant (negative).


Important Points
 CBIR has become a very active area research for two major research
communities, Database Management and Computer Vision
 Feature Extraction methods are easy, effective and less expensive
 FCH and Gabor wavelet transform are found to yield the highest color and
texture retrieval results, respectively, at the cost of higher computational
complexity.
 To make the content-based Image Retrieval truly scalable to large size
image collections, multi-dimensional indexing techniques are highly
essential.

Content based image retrieval

  • 1.
    Evaluation of Content-BasedImage Retrieval Systems Based on Feature Extraction Presented by Fouzia Ashraf Khan
  • 2.
    Motivation  Exponential increasein computing power and electronic storage capacity  Exponential increase in digital image/video database sizes  Increase use of image and video: Entertainment Education Commercial purposes  Need abstractions for efficient and effective modeling  Continuous data, multiple features and analysis of multi-layer or multivariate data processing is difficult job.
  • 3.
    Spatial Database  SpatialData  Displayed, manipulated and analyzed by means of spatial attribute which denotes a location on or near the surface of earth Spatial attribute is presented in the form of co-ordinate pairs that allow position and shape of a particular feature to be measured and represented graphically. • Spatial Database A spatial database is a database that is optimized to store and query data related to objects in space, including points, lines and polygons Typical databases understand various numeric and character data Need additional functionality spatial data types - called geometry or feature.
  • 4.
    CBIR • Content-Based ImageRetrieval - a technique that utilizes the visual content of an image, to search for similar images in large-scale image databases, according to a user’s interest - is a robust system based upon a combination of higher- level and lower-level vision principles. - Higher-level analysis uses perceptual organization, inference and grouping principles to extract semantic information - Lower-level analysis employs a channel energy model to describe image texture, shape and utilizes color histogram techniques
  • 5.
    CBIR System Components Three Major Components  Feature Extraction  Indexing  System Design Challenge: Gap between low-level features and high level user semantics
  • 6.
    Feature Extraction  Feature Acluster of points or a boundary /region of pixels  Primary Features ◦ Color ◦ Texture ◦ Shape ◦ Spatial location • Feature extraction Process of studying and locating areas and objects on the ground and deriving useful information from images.  Feature Extraction Methods ◦ Fuzzy approach ◦ Relevance feedback (supervised learning) Example of a Classified Image
  • 7.
    Color Based FeatureExtraction Approaches  Conventional Color Histogram (CCH)  Frequency of occurrence of every color in image.  Refers probability mass function  Captures the joint probabilities of intensities of color channels.  hA,B,C(a,b,c) = N. Prob(A=a, B=b, C=c), where A, B and C are the three color channels and N is the number of pixels in the image  Fuzzy Color Histogram (FCH)  Considers degree of color similarity between pixels  A color space with K color bins, FCH of I is defined as F(I)=[f1,f2,…fk] where fi = and where N = # of pixels μij = Membership value of jth pixel to ith color bin dij = Euclidean distance between the color of pixel jth and the ith color bin, ς = Average distance between colors in quantized color space  N i ij N )( 1    dijxfij   1 1 )(
  • 8.
    Color Based FeatureExtraction Approaches  Color Correlogram (CC)  Express how spatial correlation of pairs of colors changes with distance.  A table indexed by color pairs  For dth entry at location (i,j) is computed by - Counting number of pixels of color j at a distance d from a pixel of color i, divided by the total number of pixels in the image  Color-Shape Based Method  Includes area and shape information  Area of each object is encoded as the number of pixels.  Shape is characterized by “perimeter intercepted lengths”  PIL - obtained by intercepting object perimeter with 8 line segments having 8 different orientations and passing through object center.
  • 9.
    Comparative Study ofColor Based Feature Extraction
  • 10.
    Texture Based FeatureExtraction Approaches  Steerable Pyramid ◦ Basic filters are translation and rotation of a single function ◦ Filter is linear combination of basis functions ◦ Only for rotation-invariant texture retrieval  Contourlet Transform ◦ Combination of a Laplacian pyramid and directional filter bank ◦ LP provides the multi‐scale decomposition, ◦ DFB provides the multi‐directional decomposition  Gabor Wavelet ◦ Optimally achieves joint resolution in space and spatial frequency ◦ Transform dilates and rotates the two‐dimensional Gabor function. ◦ Image is convolved with each of the obtained Gabor functions ◦ Gabor function works in Fourier domain, ◦ Computationally intensive  Complex Directional Filter Bank (CDFB) ◦ Retrieval results comparable with Gabor wavelet results ◦ Shift Invariant
  • 11.
    Comparative Study ofTexture Based Feature Extraction
  • 12.
    Shape Based Feature ExtractionApproaches  Blob Detection  Detecting points , regions are either brighter or darker than surrounding  Two classes of blob detectors - Differential method based on derivative expressions - Methods based on local extrema of intensity • Thresholding  Classify pixels into two categories: – Those to which some property measured from the image falls below a threshold, and those at which the property equals or exceeds a threshold. – Thresholding creates a binary image : binarization - Choosing a threshold is a critical task.
  • 13.
     Hough Transform -Find imperfect instances of objects within a certain class of shapes by a voting procedure - Voting procedure is carried out in a parameter space - Object candidates are obtained as local maxima  Template Matching -For finding small parts of an image which match a template -To detect edges in images. -Simple to implement and understand, it is one of the slowest methods.
  • 14.
    High Dimensional Indexing Techniques To make CBIR truly scalable to large size image collections  Some Multi-dimensional Indexing Methods  Self-Organizing Maps - SOM are feed-forward, competitive artificial neural networks. - Reduces complexity (both time and computation)  Shape-Depth Representation -Well-defined geometric objects, a novel strategy based on a 3D embedding -Is between contour-based and region-based representations -A very useful one for retrieving objects/images  Fourier Transform of Shape-Depth and Indexing - Invariant to transformations such as translation, scaling and rotation. - Fourier transform is widely used for achieving the invariance
  • 15.
    System Design  Somefactors of Design - Random browsing - Search by example - Search by sketch - Search by text (including keyword or speech) and navigation with customized image categories. Some Current Developed CBIR Systems QBIC - Query By Image Content Virage RetrievalWare VisualSEEk and WebSEEk Netra
  • 16.
    Performance Evaluation ofCBIR System Performance  Observed Probability  Distance Measure and Retrieval Score  measure - user is facing a target search task from a database D for an image I belonging to class C D. - Resembles the “target testing" method - Obtained by implementing an “ideal screener“ - A computer program which simulates the human user by examining the output of the retrieval system - Marking the images returned by the system either as relevant (positive) or non-relevant (negative).  
  • 17.
    Important Points  CBIRhas become a very active area research for two major research communities, Database Management and Computer Vision  Feature Extraction methods are easy, effective and less expensive  FCH and Gabor wavelet transform are found to yield the highest color and texture retrieval results, respectively, at the cost of higher computational complexity.  To make the content-based Image Retrieval truly scalable to large size image collections, multi-dimensional indexing techniques are highly essential.