On the Utility of Input Selection and Pruning for

Download Report

Transcript On the Utility of Input Selection and Pruning for

Content-Based Image Retrieval
Using Grey Relational Analysis
Dept. of Computer Engineering
Tatung University
Presenter: Tienwei Tsai (蔡殿偉)
2005/12/02
1
Outline
Introduction
Feature Extraction
Color Space Transformation
Discrete Cosine Transform (DCT)
Similarity Measurement
Problem Description
Grey Relational Analysis
Grey Similarity Measurement
Experimental Results
Conclusions
2005/12/02
2
Introduction (1/4)
Query-by-text has two major drawbacks
 Each image in the database has to be described by
keywords, which is extremely time consuming.
 The expressive power of keywords is limited and
cannot be exhaustive.
Retrieving images according to the image
contents is called content-based image retrieval
(CBIR).
Image objects are typically represented in CBIR
 Low-level features including color, texture, and
shape, etc.
 Usually stored in the form of a feature vector.
2005/12/02
3
Introduction (2/4)
Searching for k most-similar images to a query
image
 Comparing the feature vectors of all the images in
the database with that of the query image using
some pre-selected similarity measure.
 This requires linear time with respect to the size of
the database and quickly becomes impractical for
large databases.
 A compact representation of the features needs to
be carefully chosen before applying any of these
indexing techniques
2005/12/02
4
Introduction (3/4)
We intend to achieve an acceptable retrieval
result using
 A content descriptor formed by only one single
feature, the Y-component in YUV color space.
 The first 16 DCT coefficients transformed from the Ycomponent of an image.
The run-time complexity of this method is small,
since the length of the feature vector is small.
2005/12/02
5
Introduction (4/4)
While retrieving images, the query results are
uncertain to some extent; therefore, retrieval
process can be regarded as a grey system.
The Gray Relational Analysis (GRA) method
 Describes the relationship between a main factor and
all the other factors in an uncertainly and incompletely
grey environment.
We develop a CBIR matching technique based on
the GRA method.
This method performs a high efficiency of
retrieval with an acceptable accuracy.
2005/12/02
6
Feature Extraction (1/5)
CBIR often comprises both indexing and
retrieval.
Feature extraction finds out the suitable
properties of interest and converting them into
mathematical “feature vectors”.
The feature vectors are stored into the database
along with original images.
It is a critical phase in CBIR, because the
following retrieval process depends upon the
correctness of the selected features.
2005/12/02
7
Feature Extraction (2/5)
DCT is one of the best filters for feature
extraction in the spatial frequency domain.
The steps of the proposed feature extraction
 Transform a RGB image into the YUV color space.
 Calculate the low frequency DCT coefficients for Ycomponent.
2005/12/02
8
Feature Extraction (3/5)
YUV color space
 Y: luminance (or brightness)
Y(x, y) = 0.299 R(x, y)+0.587 G(x, y)+0.114 B(x, y)
 U: blue chrominance
U(x, y) = 0.492(B(x, y)-Y(x, y))
 V: red chrominance
V(x, y) = 0.877(R(x, y)-Y(x, y))
Psycho-perceptual studies have shown that the
human brain perceives images largely based on
their luminance value, and only secondarily
based on their color information
2005/12/02
9
Feature Extraction (4/5)
Discrete Cosine Transform
 DCT coefficients are generated for a N×N image on a
pixel by pixel basis. The N×N DCT coefficients thus
give the nature of textual energy for each pixel.
N 1 N 1
2
(2i  1)u
(2 j  1)v
C (u, v)   (u ) (v) x(i, j )  cos(
) cos(
),
N
2N
2N
i 0 j 0
where x(i, j) is the pixel value at the (i, j) coordinate
position in the image, C(u, v) is DCT domain
representation of x(i, j), and
 1
 ( w)   2
 1
2005/12/02
for w  0,
otherwise.
10
Feature Extraction (5/5)
The coefficients with small u and v correspond
to low frequency components.
For most images, much of the signal energy lies
at low frequencies.
We generate a feature vector formed with low
frequency DCT terms transformed from the Ycomponent of an image .
Our experiments have shown that the use of a
block size of 4x4 low frequency coefficients
performs well from a viewpoint of retrieval
quality.
2005/12/02
11
Similarity Measurement (1/5)
The user query feature extraction will first
convert a user's exemplary image into a
mathematical feature vector compatible with the
feature vectors stored in the database.
Similarity matching is typically defined by a
distance function, e.g. Euclidean distance.
 Fewer items in a feature vector will lead to a faster
matching at the expense of accuracy.
2005/12/02
12
Similarity Measurement (2/5)
Grey system theory can perform grey relational
analysis (GRA) for data sequences.
GRA can be used to determine the relational
grade between the reference and each sequence
in the given set.
 The best comparative one can be found by further
sorting the resultant relational grades .
2005/12/02
13
Similarity Measurement (3/5)
The GRA method can be described as follows.
 Suppose the reference sequence is
y0  ( y0 (1), y0 (2),..., y0 (n)),
where n stands for the number of items in a data
sequence. Denote the m sequences to be compared
by
xi  ( xi (1), xi (2),...,xi (n)), wherer i  1,..., m.
 Define the grey relational coefficient between y0 and
xi at the kth item as follows:
2005/12/02
14
Similarity Measurement (4/5)
 ( y0 (k ), xi (k )) 
where
 min     max
,
 0i (k )     max
0i (k )  y0 (k )  xi (k )
 min  min min  0i (k ),
i
k
 max  max max  0i (k ),
i
k
i = 1…, m, k = 1, …, n, and   (0,1].
 is a distinguishing coefficient to control the resolution
between  min and max
It is set to 0.5 in this study.
2005/12/02
15
Similarity Measurement (5/5)
The grey relational grade for an entire sequence
1 n
 ( y0 , xi )   ( y0 (k ), xi (k )).
n k 1
 ( y0 , xi )
represents to the degree of similarity between the
sequence xi and the reference sequence y0.
The larger the grey relational grade  ( y0 , xi ) is,
the higher the relative similarity of xi to y0.
2005/12/02
16
Experimental Results
Two image databases are used in the system:
 1000 images downloaded from the WBIIS.
 1000 images downloaded from the test database
used in SIMPLIcity paper.
No pre-processing was done on the images.
We conducted three queries using the proposed
GRA method in this system.
There had been very little systematic evaluation
of CBIR, particularly in situations in which the
user does not have a target image in mind. .
2005/12/02
17
Figure 1. The main screen of the proposed
system
2005/12/02
18
The first query
Figure 2. Retrieved results using the GRA method.
An image of a flower was given in the first query. We
obtained eight correct answers (related to flowers) in
the top 10 images.
2005/12/02
19
The second query
Figure 3. Retrieved results using the GRA method.
An image of the sunset was given in the second query.
We obtained three correct answers (related to
sunsets) in the top 10 images.
2005/12/02
20
The third query
Figure 4. Retrieved results using the GRA method.
An image of some people was given in the third query.
We obtained nine correct answers (related to people)
in the top 10 images.
2005/12/02
21
Conclusions (1/2)
A CBIR method based on the GRA method is
proposed.
Each image is first transformed from the
standard RGB color space to the YUV space;
then Y component of the image is further
transformed to its DCT domain.
The low frequency DCT coefficients are applied
to extract low-level features from the images
due to its superiority in energy compacting.
2005/12/02
22
Conclusions (2/2)
The retrieval process is treated as a grey system.
We define a measure using GRA method to
indicate the degree of similarity between the
query image and candidate images in the
database.
Those images with the largest value of grey
relational grade are chosen as the query results.
The experimental system demonstrated the
efficiency and effectiveness of our proposed
method.
2005/12/02
23