Content Based Image Retrieval Approach Based on Top-Hat Transform And Modified Local Binary Patterns

Sundarapandian et al. (Eds): ICAITA, SAI, SEAS, CDKP, CMCA, CS & IT 08,
pp. 59–70, 2012. © CS & IT-CSCP 2012 DOI : 10.5121/csit.2012.2506
Content Based Image Retrieval Approach Based
on Top-Hat Transform And Modified Local
Binary Patterns
Mohammad Saberi1
, Farshad Tajeripour2
, and Shervan Fekri-Ershad3
1,2,3
Department of Electrical and Computer Engineering,
Shiraz University, Shiraz, Iran
1
m_saberi@shirazu.ac.ir 2
tajeri@shirazu.ac.ir
3
shfekri@shirazu.ac.ir
ABSTRACT
In this paper a robust approach is proposed for content based image retrieval (CBIR) using
texture analysis techniques. The proposed approach includes three main steps. In the first one,
shape detection is done based on Top-Hat transform to detect and crop object part of the image.
Second step is included a texture feature representation algorithm using color local binary
patterns (CLBP) and local variance features. Finally, to retrieve mostly closing matching
images to the query, log likelihood ratio is used. The performance of the proposed approach is
evaluated using Corel and Simplicity image sets and it compared by some of other well-known
approaches in terms of precision and recall which shows the superiority of the proposed
approach. Low noise sensitivity, rotation invariant, shift invariant, gray scale invariant and low
computational complexity are some of other advantages.
KEYWORDS
Image retrieval, Local binary pattern, Local variance, Top-Hat Transform, Texture analysis
1. INTRODUCTION
The requirement to find the relevant images from huge databases created due to recent
technological developments in diversity of domains is an important field of research which calls
image retrieval. The need to find a desired image from a collection is shared by many
professional groups, including journalists, forensic exports, design engineers, forensic experts and
art historiographers.
Content-based image retrieval (CBIR) is a technique to retrieve images on the basis of image
specific features such as color, texture and shape. In CBIR, initially, features are computed for
both stored and query images, and used to declare images most closely matching the query.
Color-based retrieval first evolved from simple statistical measures such as average color to color
histograms and spatial color descriptors [1]. Unfortunately, color histograms have limited
discriminative power with respect to the spatial organization of colors. The other popular color
features in content-based image retrieval applications include color correlograms [2], color

60 Computer Science & Information Technology (CS & IT)
moments [3] and MPEG-7 color descriptors [4]. The color correlogram describes the probability
of finding color pairs at fixed pixel distances and provide an efficient spatial retrieval feature
Partly due to the imprecise understanding and definition of what exactly visual texture actually is,
texture features have an even larger variety than color features. As for the texture feature
exactors, many approaches have been proposed based on Gabor filtering [5] Micro-Structure [6],
GLCM [7], and Primitive texture features [8].
These texture features try to capture the characteristics of the image or image parts with respect to
changes in certain directions and the scale of the changes. This is most useful for regions or
images with homogeneous texture. Again, invariances with respect to rotations of the image,
shifts or scale changes can be included into the feature space but information on the texture can
get lost in this process.
Shape is a key attribute of segmented image regions, and its efficient and robust representation
plays an important role in retrieval. Synonymous with shape representation is the way in which
such representations are matched with each other. Some researchers are used shape features to
retrieve similar images. In this respect, many techniques are used for shape detection such as
morphological operations [9], edge detection [10], primitive shape filters [11] and image
segmentation [12].
In this paper, we proposed an approach using combination of color, texture and shape features
together. It is included two main steps. First one is prepared to crop the main object part of the
query and database images based on top-hat transform as a shape detector. First step eliminates
irrelevant parts of the images to extract more valid feature vector. Second step is included an
algorithm for extracting features from object part image. Color local binary patterns (CLBP) and
local variance are used as feature exactors in this step.
According to the size of huge image databases, we proposed a novel extra algorithm to reduce
computational complexity. It is done as a preprocess step before retrieving to disregard the
absolutely not close categories to the query. The proposed preprocess algorithm compares non-
similarity amount between query image and each category to disregard nowhere near categories.
The performance of the proposed approach is evaluated by applying on Corel and Simplicity
image sets in terms of precision and recall and it is compared by some other state of the art CBIR
algorithms. Low noise sensitivity, rotation invariant, gray scale invariant, shift invariant and Low
computational complexity are some other advantages of proposed approach which are
demonstrated in the result section.
2. OBJECT CROPPING:
2.1. Top-Hat Transform
According to the image sets which are used in image retrieval problems, there are just one or
some more objects in the images and all of other regions are not relevant. The irrelevant regions
may invalid the values in extracted feature vectors. An approach to extract and crop object parts
is so necessary. Shape detection techniques detect objects in the images. In this paper an
algorithms is proposed to detect and crop object part in input images based on one of popular
shape detection techniques which calls top-hat transform.

Computer Science & Information Technology (CS & IT) 61
The groups of image processing operations which process the image based on shapes are referred
as Morphological operations. In morphological operations the output image is created with help
of applying structuring element to the input image. In a morphological operation, the value of
each pixel in the output image is based on a comparison of the corresponding pixel in the input
image with its neighbors [20].
Dilation and erosion are the most basic morphological operations. Dilation creates the swelling
effect of shapes of the objects by adding pixels to the boundaries of objects in an image, while
erosion forms the object shape shrinking effect by removing pixels on object boundaries [20].
The size and shape of structuring element decide number of pixels added or removed from the
objects in an image.
Dilation of a grey-scale image I(x, y) by a grey-scale structuring element B(s, t) is denoted by:
I B(x, y) = Max {I(x - s, y - t) + b(s, t)} (1)
The domain of I B is the dilation of the domain of "I" by the domain of "B". Erosion of a grey-
scale image I(x, y) by a grey-scale structuring element B(s, t) is denoted by:
IϴB(x, y) = min {I(x + s, y + t) - b(s, t)} (2)
The domain of IϴB is the erosion of the domain of "I" by the domain of "B".
Opening operator is erosion followed by dilation and closing operator is dilation followed by
erosion. Opening generally smoothes the contour of an image, breaks narrow gaps. As opposed to
opening, closing tends to fuse narrow breaks, eliminates small holes, and fills gaps in the
contours.
In mathematical morphology and digital image processing, top-hat transform is an operation that
extracts small elements and details from given images. There exist two types of top-hat
transform: The white top-hat transform is defined as the difference between the input image and
its opening by some structuring element (Eq. 3). The black top-hat transform is defined dually as
the difference between the closing and the input image (Eq. 4).
The white top-hat transform of input image "I" is given by:
T1= I - (I ○ S) (3)
The Black top-hat transform of input image "I" is given by:
T2= (I ● S) - I (4)
Where "I" means the input image and "S" is the structure element. "T1" shows the white top-hat
transform output and "T2" shows the black top-hat transform output. Also, "○" denotes the
opening operation and "●" denotes closing operation.
2.2. Proposed Object Cropping Algorithm
In order to detect and crop the object parts in original input image, an edge detection technique
should be used. In the previous section, black (white) top-hat transform is described as an edge

detection technique. In order to crop object parts, an algorithm is proposed in this section which
consists 3 sub-steps as follows:
I) Applying black (white) top-hat transform on input image as edge detection technique to detect
edge pixels of the input image. The output result is called "Edge-detected image".
II) In order to crop the object parts of the original input image, it’s enough to detect zone of the
object region. The zone of the object region is limited between four bands (Up, down, left and
right). The second step is detecting each of bands individually in edge-Detected image
III) Cropping each pixel in the original image which is limited in the object zone using computed
bands.
In the Fig.1, some results of applying the top-hat transform on Simplicity set images to detect and
crop object parts are shown.
(a)
(b)
(c)
(d)
Figure1. Some examples of object cropping using Corel dataset images
(a) Original of Horses category (b) Object cropped of (a), (c) Original of Flower category (d) Object
cropped of (c)
3.TEXTURE FEATURE EXTRACTION:
3.1 Local Binary Patterns
The local binary patterns (LBP) is a non-parametric operator which describes the local contrast
and local spatial structure of an image. First time [13] introduced this operator and showed its
high severability and discriminative power for texture analyzing and classification.In order to
evaluate the LBP, at a given pixel position (xc, yc), LBP is defined as an ordered set of binary
comparisons of pixel intensities between the center pixel and its surrounding pixels. Usually to
achieve the rotation invariant, neighborhoods would be assumed circular.Points which the
coordination’s are not exactly located at the center of pixel would be found by interpolation.Some
of the circular neighborhoods by different radius (R) and (P) number of neighborhoods pixels are
shown in Fig.2.
Now, the LBP are defined at a neighborhood of image by Eq.5.
‫۾۰ۺ‬‫܀,۾‬ = ෍ ф(g୩ − gୡ)2୩ିଵ
୮
୩ୀଵ
(5)

Computer Science & Information Technology (CS & IT)
Where, "gc" corresponds to the grey value of the centered pixel and "g
neighborhood pixels. So, P will be the number of neighborhoods of center pixel, and function
ф(x) is defined as:
ф(‫)ܠ‬ =
An example of applying LBP8,1
pixel is computed.
Figure2. Some examples of circular
(a) P=8 R=1 (b) P=12 R=1.5 (c) P=16 R=2
Figure
3.1.1. Modified Local Binary Patterns
Practical experience of the authors in [14], however, has shown that LBPROT as such does not
provide very good discrimination, and the computation complexity of basic LBP is high. To solve
these problems, [15] defined a uniformity measure "U", which corresponds to the number of
spatial transitions (bitwise 0/1 changes) in the "pattern". It is shown in Eq.7. Fo
patterns 00000000 and 11111111 have U value of 0, while 11001001 have U value of 3.
In this version of LBP, the patterns which have uniformity amount less than U
uniform patterns and the patterns with uniformity amount more than U
patterns. Finally, the LBP is computed by using Eq.8.
‫۾۰ۺ‬‫܀,۾‬
‫ܝܑܚ‬‫܂‬
ൌ
" corresponds to the grey value of the centered pixel and "gk" to the grey values
ሻ ቄ
1 if x ൒ 0
0 else
ሺ6ሻ
8,1 operator is shown in Fig.3. Where LBP8,1 amount of the center
(a) (b) (c)
2. Some examples of circular neighbourhoods
(a) P=8 R=1 (b) P=12 R=1.5 (c) P=16 R=2
ure3. An example of LBP8,1 texture model
3.1.1. Modified Local Binary Patterns
good discrimination, and the computation complexity of basic LBP is high. To solve
spatial transitions (bitwise 0/1 changes) in the "pattern". It is shown in Eq.7. Fo
In this version of LBP, the patterns which have uniformity amount less than UT are grouped as
uniform patterns and the patterns with uniformity amount more than UT grouped as non
patterns. Finally, the LBP is computed by using Eq.8.
(8)
63
" to the grey values of the
amount of the center
(a) (b) (c)
good discrimination, and the computation complexity of basic LBP is high. To solve
spatial transitions (bitwise 0/1 changes) in the "pattern". It is shown in Eq.7. For example,
(7)
are grouped as
grouped as non-uniform

Superscript "riuT" reflects the use of rotation invariant "unifor
most UT.
According to Eq.8, applying LBP
label "P+1" to non-uniform patterns. Because, LBP
non-uniform patterns, so uniform labels should cover mostly patterns in the image. In [16, 17]
show that if in the definition of LBP
negligible portion of the patterns in the texture takes label P+1.
3.1.2. Proposed feature extraction based on LBP
As it was described in section 3.1.1, a label is assigned to each neighborhood. Regarding the
Eq.9, if the number of neighbors is considered "P" pixels, applying LBP
from zero to "P" to uniform segments and label "P+1" to non
image, one "P+2" dimensional feature vector can be extracted.
To extract the feature vector, first the LBP
to neighbors. Then the occurrence probability of each label in the image is regarded as one of the
dimensions of the feature vector. The occurrence probability of a specific label in the image can
be approximated by the ratio of the number of that label to the number of all labels
Where, N୮ౡ
is the number of neighbors that labeled as P
neighbors. The extracted feature vector is shown in Eq.10.
(10)
Where, F is the feature vector extracted for neighbors and P
"i" in neighbors.
3.2. Local Variance
The LBPP,R operator is a gray-
monotonic transformation of the gray scale. It is an excellent measure of the spatial pattern, but it,
by definition, discards contrast [15]. If gray
incorporate the contrast of local image texture as well, we can measure it with a rotation invariant
measure of local variance (Eq. 11):
(11)
VARP,R is by definition invariant against shifts
complementary, their joint distribution LBP
invariant measure of local image texture.
" reflects the use of rotation invariant "uniform" patterns that have U value of at
According to Eq.8, applying LBPP,R will assign a label from "0" to "P" to uniform patterns and
uniform patterns. Because, LBPP,R just one label (P+1) is assigned to all of the
terns, so uniform labels should cover mostly patterns in the image. In [16, 17]
show that if in the definition of LBPP,R operator the value of UT is selected equal to (P/4), only a
negligible portion of the patterns in the texture takes label P+1.
Proposed feature extraction based on LBP
Eq.9, if the number of neighbors is considered "P" pixels, applying LBPP,R will assign a label
orm segments and label "P+1" to non-uniform segments. So, for every
image, one "P+2" dimensional feature vector can be extracted.
To extract the feature vector, first the LBPP,R is applied on the image and the labels are assigned
currence probability of each label in the image is regarded as one of the
be approximated by the ratio of the number of that label to the number of all labels
(9)
is the number of neighbors that labeled as Pk, and Ntotal is the number of all
The extracted feature vector is shown in Eq.10.
(10)
Where, F is the feature vector extracted for neighbors and Pi is the occurrence probability of label
-scale invariant measure, i.e., its output is not affected by any
by definition, discards contrast [15]. If gray-scale invariance is not required and it is wanted to
measure of local variance (Eq. 11):
ܸ‫ܴܣ‬௉,ோ ൌ
1
ܲ
෍ሺ݃௞ െ ߤሻଶ
௉
௞ୀଵ
is by definition invariant against shifts in gray scale. Since LBPP,R and VAR
complementary, their joint distribution LBPP,R /VARP,R is expected to be a very powerful rotation
invariant measure of local image texture.
m" patterns that have U value of at
will assign a label from "0" to "P" to uniform patterns and
just one label (P+1) is assigned to all of the
terns, so uniform labels should cover mostly patterns in the image. In [16, 17]
is selected equal to (P/4), only a
will assign a label
uniform segments. So, for every
is applied on the image and the labels are assigned
currence probability of each label in the image is regarded as one of the
be approximated by the ratio of the number of that label to the number of all labels (Eq.9).
is the number of all
is the occurrence probability of label
scale invariant measure, i.e., its output is not affected by any
ot required and it is wanted to
and VARP,R are
is expected to be a very powerful rotation

3.2.1. Feature extraction based on VAR
A vector showing the variance of all points in the range is calculated by applying variance
operator (VARP,R) on a neighbor in size of "P" and "R",
is divided into "K" bins and the number of variances falling in each bin is calculated.
vector shows the number of similar variances in all "K"
VARK
P,R.
The VARK
P,R is a discriminative texture feature vector which includes the local contrast
information of the input image. The results show the di
retrieving images.
4. PROPOSED IMAGE RETRIEVAL
Content-based image retrieval (CBIR) is a technique to retrieve images on the basis of image
specific features such as color, texture and shape. In CBIR,
both all of the stored and query images, and used to declare images most closely matching the
query. The general block diagram of CBIR algorithms is shown in Fig.4.
Figure4. CBIR general Block Diagram
According to the Fig.4, the main step of the CBIR algorithms is feature extraction. In the
proposed approach, the feature extraction step is included 3 sub
ones as follows:
I) Cropping the object parts of the query and database images
proposed in the section2.2. In this paper, the object cropped image is called OCI (Object
Image).
II) The OCIs are in RGB color space or may transform to the RGB space (R=red, G=green and
B=blue).
The second step is applying the LBP
extracting features based on the proposed approach in section 3.1.2 from each color channels.
3.2.1. Feature extraction based on VARP,R
e of all points in the range is calculated by applying variance
) on a neighbor in size of "P" and "R", for all points. After sorting this vector, it
bins and the number of variances falling in each bin is calculated.
vector shows the number of similar variances in all "K" bins. This vector is represented as
is a discriminative texture feature vector which includes the local contrast
information of the input image. The results show the discrimination strength of this operator in
ETRIEVAL APPROACH
based image retrieval (CBIR) is a technique to retrieve images on the basis of image
specific features such as color, texture and shape. In CBIR, initially, features are computed for
query. The general block diagram of CBIR algorithms is shown in Fig.4.
Figure4. CBIR general Block Diagram
the Fig.4, the main step of the CBIR algorithms is feature extraction. In the
proposed approach, the feature extraction step is included 3 sub-steps which is described ones by
I) Cropping the object parts of the query and database images using the algorithm which is
proposed in the section2.2. In this paper, the object cropped image is called OCI (Object
pplying the LBPP,R on each color channels of the OCIs individually and
65
e of all points in the range is calculated by applying variance
for all points. After sorting this vector, it
bins and the number of variances falling in each bin is calculated. The final
bins. This vector is represented as
is a discriminative texture feature vector which includes the local contrast
scrimination strength of this operator in
based image retrieval (CBIR) is a technique to retrieve images on the basis of image
initially, features are computed for
the Fig.4, the main step of the CBIR algorithms is feature extraction. In the
steps which is described ones by
using the algorithm which is
proposed in the section2.2. In this paper, the object cropped image is called OCI (Object-cropped
on each color channels of the OCIs individually and

After considering color features, to provide a main discriminated feature vector for each
Eq.(12) should be used which is described as follows:
‫ܨ‬஼௅஻௉
௜
ൌ ሺ‫ܨ‬௥௘ௗ
௜
൅ ‫ܨ‬௚௥௘௘௡
௜
൅ ‫ܨ‬௕௟௨௘
௜
ሻ
Where, FCLBP is the output feature vector and F
channel "d". Also, "i" is the dimension number of the feature vectors.
III) Applying the VARP,R on OCIs and extracting features based on the proposed approach in
section 3.2.1. The extracted feature vector is called F
As it was mentioned in this section, the image analy
considering color, texture and shape features together. Using the proposed feature extraction
algorithm, two feature vectors are provided for each input images either query or database
images. The total block-diagram of the proposed feature extraction algorithm is shown in the
Fig.5.
Figure5. The block-Diagram of the proposed feature extraction algorithm
According to the Fig.4, after feature extraction another important step is similarity comparing, to
retrieve database images which are closer to the query.
The extracted feature vectors are probabilistic and sum of all dimension values are one. In [17],
Tajeripour and Fekri-Ershad are used log
extracted local binary patterns feature vectors. In this respect, after researching about distance
and similarity criteria, the log-
comparing step [15].
Log-likelihood ratio between two probabilist
the Eq.(13).
Where, L is the log-likelihood ratio and K is number of dimensions of the A and B vectors.
After considering color features, to provide a main discriminated feature vector for each
Eq.(12) should be used which is described as follows:
ሻ/ 3 (12)
is the output feature vector and Fd is shown the extracted feature vector for color
is the dimension number of the feature vectors.
on OCIs and extracting features based on the proposed approach in
section 3.2.1. The extracted feature vector is called FVAR.
As it was mentioned in this section, the image analysis and feature extraction technique is done
agram of the proposed feature extraction algorithm is shown in the
Diagram of the proposed feature extraction algorithm
retrieve database images which are closer to the query.
Ershad are used log-Likelihood to evaluate non-similarity amount between
ed local binary patterns feature vectors. In this respect, after researching about distance
-likelihood ratio is selected as distance measure for similarity
likelihood ratio between two probabilistic feature vector (A, B) is computed as described in
(13)
likelihood ratio and K is number of dimensions of the A and B vectors.
After considering color features, to provide a main discriminated feature vector for each OCI, the
is shown the extracted feature vector for color
on OCIs and extracting features based on the proposed approach in
sis and feature extraction technique is done
agram of the proposed feature extraction algorithm is shown in the
Diagram of the proposed feature extraction algorithm
similarity amount between
ed local binary patterns feature vectors. In this respect, after researching about distance
likelihood ratio is selected as distance measure for similarity
ic feature vector (A, B) is computed as described in
likelihood ratio and K is number of dimensions of the A and B vectors.

To declare database images most closely matching the query, the log-likelihood ratio is computed
between query and each database image extracted feature vectors. Finally, the total similarity
amount is considered sum of two computed ratios. It is shown in the Eq. (14).
L (Q, I) = L (FQ.CBLP, FI.CBLP) + L (FQ.VAR, FI.VAR) (14)
Where, L (Q, I) is shown the similarity amount between query image (Q) and database image (I).
FCLBP and FVAR are two extracted feature vectors which are evaluated based on LBPP,R and
VARP,R.
5. RESULTS
The performance of the image retrieval approach was tested upon the following two popular
databases: (a) Simplicity images and (b) COREL 5,000 miscellaneous database which are
downloaded from [21].
The Corel image database contains a large amount of images of various contents ranging from
animals and outdoor sports to natural scenarios. The Corel contains 50 categories. There are 5000
images from diverse contents such as fireworks, bark, microscopic, tile, food texture, tree, wave,
pills and stained glass. Each category contains 100 images of size192×128 in JPEG format.
The Simplicity database includes 10 different categories. These categories are titled: Beach,
Africa, Elephants, Buildings, Buses, Dinosaurs, Horses, Flowers, Food and Mountains. Each
category has absolutely 100 images, along with some images undergoing changes due to rotation,
translation, scaling, noise injection, illumination, etc. The total number of simplicity database size
is 1000.
In order to evaluate the effectiveness of the proposed approach, the Precision and Recall curves
are adopted, which are the most common criteria used for evaluating image retrieval
performance.
Precision and Recall are defined as follows:
Precision (N) = IN / N (15)
Recall (N) = IN / M (16)
Where "IN" is the number of images retrieved in the top "N" positions that are similar to the query
image, "M" is the total number of images in the database similar to the query(each category size),
and "N" is the total number of images retrieved.
In this respect, in order to evaluate the performance, N is considered 10, 20, 30 and 40 for both of
test datasets. Also, for feature extraction, different sizes was tested for CLBP and VAR, P = 8 and
K=16 provided maximum accuracy.
According to the Corel and Simplicity categories, "M" is 100. The average retrieval precision and
recall for Corel and simplicity dataset are shown in the Table 1 and 2.
In order to compare our approach, some of other previous well-know approaches [18] and [19]
were survived.

In [18], an approach is proposed for CBIR using visually significant point features. The authors
[18], were evaluated their approach performance using simplicity dataset considering N=10, 20
and 40. The precision% is computed for each category separately. The average precision of their
approach [18] is shown in the Table 3 and it was compared by our proposed approach.
As seen from Table3, our proposed approach has reported better results than the [18] method for
7 categories and total average. G. H. Liu et al are prepared an algorithm based on multi-texton
histograms [19]. They evaluated their CBIR system performance using Corel dataset and
considering N=12. The average precision and recall based on [19], are shown in the Table 4. It is
compared by proposed approach in a same situation (M= 5000, N=12). It can be seen from the
Table 3 & 4, that our method achieves much better results than [18] and [19] methods.
Table1. The average precision and recall of Proposed approach on Corel (5000s) dataset
Dataset
Retrieved Images
Precision Recall
N = 10 53.62 5.36
N = 20 51.64 10.32
N = 30 50.24 15.07
N = 40 46.12 18.44
Table2. The average precision and recall of Proposed approach on Simplicity dataset
Dataset
Retrieved Images
Precision Recall
N = 10 81.33 8.13
N = 20 76.22 15.24
N = 30 74.86 22.45
N = 40 67.22 26.88
Table3. The average precision% of proposed and [18] on Simplicity
Approach
Category
[18] Proposed
N=10 N=20 N=40 N=10 N=20 N=40
Africa 70.50 61 58 67.34 61.04 56.76
Beach 71.48 56.23 52.58 73.10 59.11 53.87
Building 72.40 63.67 56 76.48 68.72 60.64
Bus 81.45 72.77 62.67 93.23 91.47 71.83
Dinosaur 100 95 92 100 98.87 95.48
Elephant 85 77 66 80 75.92 61.5
Flower 92 83 71.2 94.58 89.44 73.5
Horses 100 95 83 92.28 89.76 79.62
Mountains 72 68 62 72.4 68.38 63.24
Food 62 57 55.2 63.96 59.49 55.78
Average 80.68 72.86 65.86 81.33 76.22 67.22

Table4. The average precision% and Recall% of proposed and [19] on Corel, N=12
Measure
Approach
Precision Recall
[19] 49.98 6.00
Proposed 53.04 6.36
As seen from Table4, our proposed approach has reported better average precision% than the [19]
and the recall averages are near same.
6. CONCLUSION
The main aim of this paper was to propose a robust approach for content based image retrieval. In
this respect, an approach is proposed which is consists two main steps. In the first one, feature
extraction is done using combination of texture, shape and color features. In the second step, log
likelihood ratio is used to declare most closely matching images from a huge database to the
query. The performance of the approach is evaluated using Corel and Simplicity image sets and it
compared some of other well-known approaches in terms of precision and recall. The results are
shown the high performance of the proposed approach to retrieve certainly same category images
by the query. Some of other advantages of the proposed approach are as follows:
I) According to the description of local binary pattern and local variance [15], the proposed
feature extraction is rotation invariant.
II) The precision and recall amount of the proposed CBIR system is higher than many other well
known CBIR algorithms.
III) In this paper, a novel feature extraction is described using combination of top-hat transforms,
local binary patterns and local variance that can be used for other image processing cases to
analyze the color images.
REFERENCES
[1] Pass G., and Zabih R., (1999) "Comparing Images Using Joint Histograms", ACM Journal of
Multimedia Systems, 7(3), pp. 234-240
[2] Ojala T., Rautiainen M., Matinmikko E., and Aittola, M., (2001), "Semantic Image Retrieval with
HSV Correlograms", In Proc. of 12th Scandinavian Conference on Image Analysis
[3] Velmurugan K., and Santhosh Baboo S., (2011), "Content-Based Image Retrieval using SURF and
Color Moments", Global Journal of Computer Science and Technology, 11(10), pp. 1-4
[4] Eidenberger, H., (2004), "statistical analysis of MPEG-7 image descriptions", ACM Multimedia
Systems journal, 10(2), pp.84-97
[5] Manjunath B.S., and Ma W.Y., (1996) "Texture features for browsing and retrieval of image data",
IEEE Transaction on pattern analysis and machine intelligence, 18, pp. 837-842.
[6] Guang-Hai L., Zuo-Yong L., Lei Z., and Yong X., (2011), "Image retrieval based on micro-structure
descriptor", Pattern Recognition, 44(9), pp. 2123-2133
[7] Jhanwar N., Chaudhuri S., and Zavidovique B., (2004), "Content based image retrieval using motif
co-occurrence matrix", Image and Vision Computing, Vol. 22, No. 14, pp.1211-1220
[8] Hsin-Chih L., Ling-Ling W., and Shi-Nine Y., (1999), "Regular-texture image retrieval based on
texture-primitive extraction", Image and Vision Computing, 17(1) , pp.51-63

[9] Kekre H.B., Mukherjee P., and Wadhwa SH., (2010), "Image Retrieval with Shape Features Extracted
using Morphological Operators with BTC", International Journal of Computer Applications, 12(3),
pp. 1-5
[10] Berman A.P., and Shapiro, L.G., (1999), "A flexible image database system for content-based
retrieval", Computer Vision and Image Understanding, 75, pp. 175-195
[11] Greenspan S., Hayit, Malik, Jitendra, and Puzicha, (2002), "Shape matching and object recognition
using shape contexts", IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(4), pp.
509-522
[12] Lifeng L., and Sclaroff S., (1999), "Automatic Deformable Shape Segmentation for Image Database
Search Applications", In Proc. of International Conference on Visual Information
[13] Ojala, T., Maeenpaeae, T., and Pietikaeinen, M., (2000), "Texture Classification by Multi Predicate
Local Binary Pattern Operators", In Proc. 15th International Conference on Pattern Recognition, 3,
pp. 951-954.
[14] Pietikäinen, M., Ojala, T., and Xu, Z., (2000), "Rotation-Invariant Texture Classification Using
Feature Distributions", Pattern Recognition, 33, pp 43-52.
[15] Ojala, T., Pietikainen, M., and Maenpaa, T., (2002), "Multi resolution gray-scale and rotation
invariant texture classification with local binary patterns", IEEE Transactions on Pattern Analysis and
Machine Intelligence, 24(7), pp. 971–987.
[16] Tajeripour, F., Kabir, E., and Sheikhi, A., (2008), "Fabric Defect detection using Modified Local
Binary Patterns", EURASIP Journal on Advances in Signal Processing, 1, pp. 1-12.
[17] Fekri-Ershad, Sh., and Tajeripour, F., (2012), "A Robust Approach for Surface Defect Detection
Based on One Dimensional Local Binary Patterns", Indian Journal of Science and Technology, 5(8),
pp.3197-3203
[18] Banerjeea, M., Malay, K.K., Maji, P., (2009), "Content-based image retrieval using visually
significant point features", Fuzzy Sets and Systems, 160, pp. 3323-3341
[19] Guang-Hai, L., Lei, Z., Ying-Kun H., Zuo-Yong, Li., Jing-Yu, Y., (2010), " Image retrieval based on
multi-texton histogram", Pattern Recognition, 43, pp. 2380-2389
[20] Chen, T., Wu, Q.H., Rahmani-Torkaman, R., Hughes, J., (2002), "A pseudo top-hat mathematical
morphological approach to edge detection in dark regions", Pattern Recognition, 35, pp. 199-210
[21] http:bergman.stanford.edu/cgi-bin
[22] Randen, T., and Husoy, J.H., (1999), "Filtering for Texture Classification: A Comparative Study",
IEEE Trans. Pattern Analysis and Machine Intelligence, 21(4), pp. 291-310
[23] Wu, J.K. and Narasimhalu A.D., (1998) "Fuzzy content based retrieval in Image database",
Information Processing and Management, 34(5), pp. 513-534,
[24] Santini, S., Jain R., (1999), "Similarity Measure", IEEE Transaction on Pattern Analysis and Machine
Intelligence, Sep, pp. 871-883

Content Based Image Retrieval Approach Based on Top-Hat Transform And Modified Local Binary Patterns

More Related Content

What's hot (19)

Similar to Content Based Image Retrieval Approach Based on Top-Hat Transform And Modified Local Binary Patterns (20)

More from cscpconf (20)

Recently uploaded (20)

Content Based Image Retrieval Approach Based on Top-Hat Transform And Modified Local Binary Patterns