(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 7, October 2010
Sector Mean with Individual Cal and SalComponents in Walsh Transform Sectors as FeatureVectors for CBIR
H. B. Kekre
Senior Professor, Computer Engineering
MPSTME,SVKM‟S NMIMS University,
Mumbai, INDIAhbkekre@yahoo.com
Dhirendra Mishra
Associate Professor, Computer Engineering
MPSTME, SVKM‟S NMIMS University,
Mumbai, INDIAdhirendra.mishra@gmail.com
Abstract-
We have introduced a novel idea of conceivingcomplex Walsh transform for sectorization of transformedcomponents. In this paper we have proposed two differentapproaches for feature vector generation with consideration of all sal components and all cal components separately. Boththese approaches are experimented with the extra componentsof zero-cal and highest-sal. Two similarity measures such assum of absolute difference and Euclidean distance are used andresults are compared. The cross over point performance of overall average of precision and recall for both approaches ondifferent sector sizes are compared. The individual sector meanof Walsh sectors in all three color planes are considered todesign the feature vector. The algorithm proposed here isworked over database of 1055 images spread over 12 differentclasses. Overall Average precision and recall is calculated forthe performance evaluation and comparison of 4, 8, 12 & 16Walsh sectors. The use of Absolute difference as similaritymeasure always gives lesser computational complexity andconsideration of only all cal components with augmentation of zero-cal approach with sum of absolute difference as similaritymeasure of feature vector has the best retrieval performance.
Index Terms-
CBIR, Walsh Transform, Euclidian Distance,Absolute Difference, Precision, Recall
I. INTRODUCTIONWith the huge growth of digital information the need of itsmanagement requires need of storage and utilization inefficient manner. This has lead to approach like contentbased image search and retrieval to be used. Content-basedimage retrieval into automatic retrieval of images from adatabase by color, texture and shape features. The term hasbeen widely used to describe the process of retrievingdesired image on the basis of features (such as colors,texture and shape) that can be automatically extracted fromthe images themselves. The typical CBIR system [1-6]performs two major tasks. The first one is feature extraction(FE), where a set of features, called image signature orfeature vector, is generated to accurately represent thecontent of each image in the database. A feature vector ismuch smaller in size than the original image, typically of theorder of hundreds of elements (rather than millions). Thesecond task is similarity measurement (SM), where adistance between the query image and each image in thedatabase using their signatures is computed so that the topclosest images can be retrieved.[7-9]. There are variousapproaches which have been experimented to generate theefficient algorithm for CBIR like FFT sectors [4-6],Transforms [15][17], Vector quantization[12], bit truncationcoding [13][14]. In this paper we have introduced a novelconcept of complex Walsh transform and its sectorizationfor feature extraction (FE).Two different similarity measuresnamely sum of absolute difference and Euclidean distanceare considered. The performances of these approaches arecompared.II.
WALSH TRANSFORMWalsh transform [17] matrix is defined as a set of N rows,denoted Wj, for j = 0, 1, .... , N - 1, which have thefollowing properties:Wj takes on the values +1 and -1.Wj[0] = 1 for all j.Wj x W
Tk
=0, for j ≠
k and Wj x Wk
T
=N, for j=k.Wj has exactly j zero crossings, for j = 0, 1, ., N-1.Each row Wj is either even or odd with respect toits midpoint.Walsh transform matrix is generated using a Hadamardmatrix of order N. The Walsh transform matrix row is therow of the Hadamard matrix specified by the Walsh codeindex, which must be an integer in the range [0, ..., N - 1].For the Walsh code index equal to an integer j, therespective Hadamard output code has exactly j zerocrossings, for j = 0, 1, ... , N - 1.
156http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 7, October 2010
Kekre‟s Algorithm
[10] to generate Walsh Transform fromHadamard matrix [17] is illustrated for N=16.However thealgorithm is general and can be used for any N = 2
k
where k is an integer.
Step 1:
Arrange
the „N‟ numbers
in a row and then split the row at
„N/2‟, the other part is written below the upper row but in
reverse order as follows:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1515 14 13 12 11 10 9 8Step 2:
We get two rows, each of t
his row is again split in „N/4‟ and
other part is written in reverse order below the upper rows asshown below.
0 1 2 315 14 13 127 6 5 48 9 10 11
This step is repeated until we get a single column whichgives the ordering of the Hadamard rows according tosequency as given below:
0,15,7,8,3,12,4,11,1,14,6,9,2,13,5,10
Step 3:
According to this sequence the Hadamard rows are arrangedto get Walsh transform matrix. Now a product of Walshmatrix and the image matrix is calculated. This matrixcontains Walsh transform of all the columns of the givenimage.Since Walsh matrix has the entries either +1 or -1 there isno multiplication involved in computing this matrix. Sinceonly additions are involved computational complexity isvery low.III.
FEATURE
VECTOR
GENERATIONThe proposed algorithm makes novel use of Walshtransform to design the sectors to generate the featurevectors for the purpose of search and retrieval of databaseimages. The complex Walsh transform is conceived bymultiplying all sal functions by j =
√
-1 and combining themwith real cal functions of the same sequency. Thus it ispossible to calculate the angle by taking tan
-1
of sal/cal.However the values of tan are periodic with the period of
π
radians hence it can resolve these values in only two sectors.To get the angle in the range of 0-360 degrees we dividethese points in four sectors as explained below. These foursectors are further divided into 8, 12 and 16 sectors. Wehave proposed two different approaches for feature vectorgeneration namely sector mean of only sal components andonly cal components value of all the vectors in each sectorwith augmentation of extra highest-sal, zero-cal componentsand without augmentation of extra highest-sal, zero-calcomponents with sum of absolute difference and Euclideandistance [7-9] [11-14] as similarity measures. Performancesof all these approaches are compared using both similaritymeasures.
A.Four Walsh Transform Sectors:
To get the angle in the range of 0-360 degrees, the steps asgiven in Table 1 are followed to separate these points intofour quadrants of the complex plane. The Walsh transformof the color image is calculated in all three R, G and Bplanes. The complex rows representing sal components of the image and the real rows representing cal componentsare checked for positive and negative signs. The sal and calWalsh values are assigned to each quadrant. as follows:T
ABLE
I.
F
OUR
W
ALSH
S
ECTOR FORMATION
Sign of Sal Sign of CalQuadrant Assigned+ +
I (0
–
90 Degrees)
+ -
II ( 90
–
180 Degrees)
- -
III( 180- 270 Degrees)
- +
IV(270
–
360 Degrees)The equation (1) is used to generate individual componentsto generate the feature vector of dimension 12 consideringthree R, G and B Planes in the sal and cal densitydistribution approach. However, it is observed that thedensity variation in 4 quadrants is very small for all theimages. Thus the feature vectors have poor discretionarypower and hence higher number of sectors such as 8, 12 and16 were tried. In the case of second approach of featurevector generation i.e. individual sector mean has betterdiscretionary power in all sectors.Sum of absolute differencemeasure is used to check the closeness of the query imagefrom the database image and precision and recall arecalculated to measure the overall performance of thealgorithm.
B. Eight Walsh Transform Sectors:
Each quadrants formed in the previous obtained 4 sectorsare individually divided into 2 sectors each considering theangle of 45 degree. In total we form 8 sectors for R,G and Bplanes separately as shown in the Table 2. The percentage
157http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 7, October 2010
density distribution of sal and cal in all 8 sectors aredetermined using equation (1) to generate the feature vector.T
ABLE
2.
E
IGHT
W
ALSH
S
ECTOR FORMATION
Quadrant of 4Walsh sectorsCondition New sectors Formed
I (0
–
90
0
) Cal >= Sal I (0-45 Degrees)Sal > Cal II (45-90 Degrees)II ( 90
–
180
0
)
|Sal | > |Cal| III(90-135 Degrees)|Cal| >= |Sal| IV(135-180 Degrees)III ( 180- 270
0
)
|Cal| >= |Sal| V (180-225 Degrees )|Sal| > |Cal| VI (225-270 Degrees)IV ( 270
–
360
0
) |Sal| > |Cal| VII (270-315Degrees)|Cal| >= |Sal| VIII (315-360Degrees )
C. Twelve Walsh Transform Sectors:
Each quadrants formed in the previous section of 4 sectorsare individually divided into 3 sectors each considering theangle of 30 degree. In total we form 12 sectors for R,G andB planes separately as shown in the Table 3. The percentagedensity distribution of sal and cal in all 12 sectors aredetermined using equation (1) to generate the feature vectorT
ABLE
3.
T
WELVE
W
ALSH
S
ECTOR FORMATION
4 Quadrants Condition New sectors
I (0
–
90
0
)
Cal >= √3
* Sal I (0-30
0
)
1/√3
cal <=sal<=
√3
calII (30-60
0
)Otherwise III (60-90
0
)II ( 90
–
180
0
)
Cal >=
√3
* Sal IV (90-120
0
)1/
√3 |
cal| <=|sal|<=
√3 |
cal|V (120-150
0
)Otherwise VI (150-180
0
)III(180-270
0
)
|Cal|>=
√3
* |Sal| VII (180-210
0
)1/
√3
cal <=|sal|<=
√3
|cal|VIII(210-240
0
)Otherwise IX (240-270
0
)IV ( 270
–
360
0
)|Cal|>=
√3
* |Sal| X (270-300
0
)1/
√3 |
cal| <=|sal|<=
√3 |
cal|XI (300-330
0
)Otherwise XII (330-360
0
)IV.
RESULTS
AND
DISCUSSIONThe sample Images of the database of 1055 images of 12different classes such as Flower, Sunset, Barbie, Tribal,Puppy, Cartoon, Elephant, Dinosaur, Bus, Parrots, Scenery,Beach are shown in the Figure 1.The algorithm is tested bytaking 5 query images from each class and then averagingthe performance in terms of precision and recall over all theclasses.Figure 1. Sample Image DatabaseFigure 2. Query Image
158http://sites.google.com/site/ijcsis/ISSN 1947-5500
Search History:
Result 00 of 00
00 results for result for
p.
Sector Mean with Individual Cal and Sal Components in Walsh Transform Sectors as Feature Vectors for CBIR
We have introduced a novel idea of conceiving complex Walsh transform for sectorization of transformed components. In this paper we have proposed two different approaches for feature vector generation with consideration of all sal components and…
(More)
247
Reads
1
Readcasts
2
Embed Views
Get Scribd Mobile
To get Scribd mobile enter your number and we'll send you a link to the Scribd app for iPhone & Android.We've sent a link to the Scribd app. If you didn't receive it, try again.
We'll never share your phone number.
More From This User
Notes
Load more
