Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Save to My Library
Look up keyword
Like this
2Activity
0 of .
Results for:
No results containing your search query
P. 1
Density Distribution and Sector Mean with Zero-cal and Highest-sal Components in Walsh Transform Sectors as Feature Vectors for Image Retrieval

Density Distribution and Sector Mean with Zero-cal and Highest-sal Components in Walsh Transform Sectors as Feature Vectors for Image Retrieval

Ratings: (0)|Views: 22 |Likes:
Published by ijcsis
We have introduced a novel idea of considering complex Walsh transform for sectorization of transformed components. In this process the first coefficient of zero-cal and the last coefficient highest-sal are not used. In this paper we have proposed two different approaches along with the extra components of zero-cal and highest-sal for feature vector generation namely sector density and sector mean. Two similarity measures such as sum of absolute difference and Euclidean distance are used and results are compared. The cross over point performance of overall average of precision and recall for both approaches on different sector sizes are compared. The density distribution of real (cal) and imaginary (sal) values and sector mean of Walsh sectors in all three color planes are considered to design the feature vector. The algorithm proposed here is worked over database of 1055 images spread over 12 different classes. Overall Average precision and recall is calculated for the performance evaluation and comparison of 4, 8, 12 & 16 Walsh sectors. The use of sum of absolute difference as similarity measure always gives lesser computational complexity and density distribution approach with sum of absolute difference as similarity measure of feature vector has the best retrieval performance.
We have introduced a novel idea of considering complex Walsh transform for sectorization of transformed components. In this process the first coefficient of zero-cal and the last coefficient highest-sal are not used. In this paper we have proposed two different approaches along with the extra components of zero-cal and highest-sal for feature vector generation namely sector density and sector mean. Two similarity measures such as sum of absolute difference and Euclidean distance are used and results are compared. The cross over point performance of overall average of precision and recall for both approaches on different sector sizes are compared. The density distribution of real (cal) and imaginary (sal) values and sector mean of Walsh sectors in all three color planes are considered to design the feature vector. The algorithm proposed here is worked over database of 1055 images spread over 12 different classes. Overall Average precision and recall is calculated for the performance evaluation and comparison of 4, 8, 12 & 16 Walsh sectors. The use of sum of absolute difference as similarity measure always gives lesser computational complexity and density distribution approach with sum of absolute difference as similarity measure of feature vector has the best retrieval performance.

More info:

Published by: ijcsis on Aug 13, 2010
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

09/09/2010

pdf

text

original

 
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 08, No. 04, 2010
Density Distribution and Sector Mean with Zero-cal and Highest-sal Components in WalshTransform Sectors as Feature Vectors for ImageRetrieval
H.B.Kekre
Sr. ProfessorMPSTME, SVKM’s NMIMS (Deemed-to be-University)Vile Parle West, Mumbai -56,INDIAhbkekre@yahoo.com
Dhirendra Mishra
Assistant Professor & PhD Research ScholarMPSTME, SVKM’s NMIMS (Deemed-to be-University)Vile Parle West, Mumbai -56,INDIAdhirendra.mishra@gmail.com
 Abstract-
 
We have introduced a novel idea of consideringcomplex Walsh transform for sectorization of transformedcomponents. In this process the first coefficient of zero-caland the last coefficient highest-sal are not used. In this paperwe have proposed two different approaches along with theextra components of zero-cal and highest-sal for featurevector generation namely sector density and sector mean.Two similarity measures such as sum of absolute differenceand Euclidean distance are used and results are compared.The cross over point performance of overall average of precision and recall for both approaches on different sectorsizes are compared. The density distribution of real (cal) andimaginary (sal) values and sector mean of Walsh sectors in allthree color planes are considered to design the feature vector.The algorithm proposed here is worked over database of 1055images spread over 12 different classes. Overall Averageprecision and recall is calculated for the performanceevaluation and comparison of 4, 8, 12 & 16 Walsh sectors. Theuse of sum of absolute difference as similarity measure alwaysgives lesser computational complexity and density distributionapproach with sum of absolute difference as similaritymeasure of feature vector has the best retrieval performance.
 Keywords-CBIR, Walsh Transform, Euclidian Distance, Absolute Difference, Precision, Recall 
I. INTRODUCTION
With the huge growth of digital information the need of itsmanagement requires need of storage and utilization in efficientmanner. This has lead to approach like content based imagesearch and retrieval to be used. The earliest use of the termcontent-based image retrieval in the literature seems to have beenby Kato [1], to describe his experiments into automatic retrievalof images from a database by color, texture and shape features.The term has since been widely used to describe the process of retrieving desired images from a large collection on the basis of features (such as colors, texture and shape) that can beautomatically extracted from the images themselves. The typicalCBIR system [1-6] performs two major tasks. The first one isfeature extraction (FE), where a set of features, called imagesignature or feature vector, is generated to accurately representthe content 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). The secondtask is similarity measurement (SM), where a distance betweenthe query image and each image in the database using theirsignatures is computed so that the top closest images can beretrieved.[7-9]. There are various approaches which have beenexperimented to generate the efficient algorithm for CBIR likeFFT sectors [4-6], Transforms [15][17], Vector quantization[12],bit truncation coding [13][14]. In this paper we have introduced anovel concept of complex Walsh transform and its sectorizationfor feature extraction (FE).Two different similarity measuresnamely sum of absolute difference and Euclidean distance areconsidered. The performances of these approaches are compared.
II.
 
WALSH TRANSFORM
Walsh transform [17] matrix is defined as a set of N rows,denoted Wj, for j = 0, 1, .... , N - 1, which have the followingproperties:
 
Wj takes on the values +1 and -1.
 
Wj[0] = 1 for all j.
 
Wj x W
T
=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 Hadamard matrix of order N. The Walsh transform matrix row is the row of theHadamard matrix specified by the Walsh code index, which mustbe an integer in the range [0, ..., N - 1]. For the Walsh code indexequal to an integer j, the respective Hadamard output code hasexactly j zero crossings, for j = 0, 1, ... , N - 1.
341http://sites.google.com/site/ijcsis/ISSN 1947-5500
 
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 08, No. 04, 2010
Kekre’s Algorithm to generate Walsh Transform from Hadamardmatrix [17] is illustrated for N=16.However the algorithm isgeneral and can be used for any N = 2
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 orderas 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 this row is again split in ‘N/4’ and otherpart is written in reverse order below the upper rows as shownbelow.
0 1 2 315 14 13 127 6 5 48 9 10 11
This step is repeated until we get a single column which gives theordering of the Hadamard rows according to sequency as givenbelow:
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 arranged to getWalsh transform matrix. Now a product of Walsh matrix and theimage matrix is calculated. This matrix contains Walsh transformof all the columns of the given image.Since Walsh matrix has the entries either +1 or -1 there is nomultiplication involved in computing this matrix. Since onlyadditions are involved computational complexity is very low.
III.
 
FEATURE
 
VECTOR
 
GENERATION
The proposed algorithm makes novel use of Walsh transform todesign the sectors to generate the feature vectors for the purposeof search and retrieval of database images. The complex Walshtransform is conceived by multiplying all sal functions by j =
-1and combining them with real cal functions of the same sequency.Thus it is possible to calculate the angle by taking tan
-1
of sal/cal.However the values of tan are periodic with the period of 
п
 radians hence it can resolve these values in only two sectors. Toget the angle in the range of 0-360 degrees we divide these pointsin four sectors as explained below. These four sectors are furtherdivided into 8, 12 and 16 sectors. We have proposed two differentapproaches for feature vector generation namely density andmean value of all the vectors in each sector with sum of absolutedifference and Euclidean distance [7-9] [11-14] as similaritymeasures. Performance of both these approaches are comparedusing both similarity measures. Each Walsh sector is representedby single percentage value of sal-cal distribution in particularsector w.r.t. all sectors for feature vector generation using formulashown in (1) and two extra components of zero-cal and highest-salare added to the feature vector.(Total number of sal in particular sector)----------------------------------------------------- X 100 (1)(Total number of sal in all sectors)Thus for 4, 8, 12 & 16 Walsh sectors 4, 8, 12 and 16 featurecomponents along with extra components of zero-cal and highest-sal for each color planes i.e. R, G and B are generated. Thus allfeature vectors are of dimension 18, 30, 42 and 54 components. Inthe second approach mean of all sal and cal of each sector andwith extra components of zero-cal and highest-sal is calculatedand equation (2) is used for feature component representing eachsector for all three color planes i.e. R. G and B with single valuethus forming the feature vector of dimension 18, 30, 42 and 54 for4, 8, 12 and 16 complex Walsh transform sectors.
 A.
Four Walsh Transform Sectors:
To get the angle in the range of 0-360 degrees, the steps as givenin Table 1 are followed to separate these points into fourquadrants of the complex plane. The Walsh transform of the colorimage is calculated in all three R, G and B planes. The complexrows representing sal components of the image and the real rowsrepresenting cal components are checked for positive andnegative signs. The sal and cal Walsh values are assigned to eachquadrant. as follows:T
ABLE
I.
 
F
OUR
W
ALSH
S
ECTOR FORMATION
 
Sign of SalSign 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 components togenerate the feature vector of dimension 12 considering three R,G and B Planes in the sal and cal density distribution approach.However, it is observed that the density variation in 4 quadrantsis very small for all the images. Thus the feature vectors havepoor discretionary power and hence higher number of sectors suchas 8, 12 and 16 were tried. In the case of second approach of feature vector generation i.e. individual sector mean has betterdiscretionary power in all sectors and equation (2) is used togenerate the individual sector components.
__________________________________
(mean of sal vector)
2
+ (mean of cal vector)
2
(2)
 
Sum of absolute difference measure is used to check thecloseness of the query image from the database image and
342http://sites.google.com/site/ijcsis/ISSN 1947-5500
 
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 08, No. 04, 2010
precision and recall are calculated to measure the overallperformance of the algorithm.
 B. Eight Walsh Transform Sectors:
Each quadrants formed in the previous obtained 4 sectors areindividually divided into 2 sectors each considering the angle of 45 degree. In total we form 8 sectors for R,G and B planesseparately as shown in the Table 2. The percentage densitydistribution of sal and cal in all 8 sectors are determined usingequation (1) to generate the feature vector.
T
ABLE
2.
 
E
IGHT
W
ALSH
S
ECTOR FORMATION
 
Quadrant of 4
 
Walsh 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 (270-315 Degrees)|Cal| >= |Sal| VIII (315-360 Degrees )
C. Twelve Walsh Transform Sectors:
Each quadrants formed in the previous section of 4 sectors areindividually divided into 3 sectors each considering the angle of 30 degree. In total we form 12 sectors for R,G and B planesseparately as shown in the Table 3. The percentage densitydistribution of sal and cal in all 12 sectors are determined using equation (1) to generate the feature vector
T
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 cal II (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
 
DISCUSSION
The sample Images of the database of 1055 images of 12 differentclasses such as Flower, Sunset, Barbie, Tribal, Puppy, Cartoon,Elephant, Dinosaur, Bus, Parrots, Scenery,Beach is shown in theFigure 1.
Figure 1. Sample Image DatabaseFigure 2. Query Image
343http://sites.google.com/site/ijcsis/ISSN 1947-5500

You're Reading a Free Preview

Download
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->