(IJCSIS) International Journal of Computer Science and Information Security,Vol. 09, No.02, 2011
Improved Shape Content Based Image RetrievalUsing Multilevel Block Truncation Coding
Dr. H.B.Kekre
1
, Sudeep D. Thepade
2
, Miti Kakaiya
3
, Priyadarshini Mukherjee
3
,Satyajit Singh
3
,Shobhit Wadhwa
31
Senior Professor,
2
Ph.D.Research Scholar & Associate Professor,
3
B.Tech StudentComputer Engineering Department, MPSTME,
SVKM’s NMIMS (Deemed
-to-be University)Mumbai, India
1
hbkekre@yahoo.com,
2
sudeepthepade@gmail.com,
3
miti.kakaiya@gmail.com,
3
muk_priyam@hotmail.com,
3
singh.satyajit1@gmail.com,
3
shobhiitwadhwa@gmail.com
Abstract
This paper presents improved content based imageretrieval (CBIR) techniques based on multilevel BlockTruncation Coding (BTC) using multiple threshold values. BlockTruncation Coding based feature is one of the CBIR methodsproposed using shape features of image. The shape averagingmethods used here are BTC Level
–
1, BTC Level
–
2, BTC Level
–
3 and BTC Level
–
4. Here the feature vector size per image isgreatly reduced by using mean of each plane and finding out thethreshold value. Then divide each plane using the thresholdvalue. In order to find out the performance of the algorithm,shape averaging is applied to calculate precision and recallvalues. Instead of using all pixel data of image as feature vectorfor image retrieval these six, twelve, twenty
–
four and forty
–
eight feature vectors for BTC Level
–
1, Level
–
2, Level
–
3 andLevel
–
4 respectively, can be used. This results in betterperformance. The proposed CBIR techniques are tested ongeneric image database having 1000 images spread across 11categories. For each proposed CBIR technique 55 queries (5 percategory) are fired on the generic image database To compare theperformance of image retrieval techniques average precision andrecall are computed of all queries. The results have shown theperformance improvement (higher precision and recall values)with proposed methods compared to BTC Level-1.Keywords- Content Based Image Retrieval (CBIR), BTC Level-1,BTC Level-2, BTC Level-3, BTC Level - 4.
I.
I
NTRODUCTION
Information retrieval (IR) is the science of searching fordocuments, for information within documents, and formetadata about documents, as well as that of searchingrelational databases and the World Wide Web. There is overlapin the usage of the terms data retrieval, document retrieval,information retrieval, and text retrieval, but each also has itsown body of literature, theory and technologies. IR isinterdisciplinary, based on computer science, mathematics,cognitive psychology, linguistics, statistics, and physics.Automated information retrieval systems are used to reducewhat has been called "information overload". Many universitiesand public libraries use IR systems to provide access to booksand journals. Web search engines are the most visible IRapplications. Images do have giant share in this informationbeing stored and retrieved.
A.
Image Retrieval
Image search is a specialized data search used to findimages. User may give a keyword, sketch or an image to imagesearch engine for retrieving the relatively similar images fromthe image databases. The similarity used for search criteriacould be meta tags, color distribution in images andregion/shape attributes. Most traditional methods of imageretrieval utilize some method of adding metadata such ascaptioning, keywords, or descriptions to the images so thatretrieval can be performed over the annotation words[23]. Thelimitations of text-based approach are that it is subject tohuman perception and the problem of annotation of images.Annotating every image is a cumbersome and expensive task.
B.
Content-based image retrieval
Content-based image retrieval (CBIR) is the application of computer vision to the image retrieval problem, that is, theproblem of searching for digital images in large databases. Theterm 'content' in this context might refer to color, shapes andtextures. The color aspect can be achieved by the techniquesaveraging and histograms [4, 5, 7]. The texture aspect can beachieved by using transforms [12] or vector quantization [9,11, 15]. Finally the shape aspect can be achieved by usinggradient operators or morphological operators. Some of themajor areas of application are: Art collections, Medicaldiagnosis, Crime prevention, the military, Intellectualproperty, Architectural and engineering design andGeographical information and remote sensing systems.II.
EDGE EXTRACTION
Edge detection is very important in image analysis. Theedges give idea about the shapes of objects present in theimage. Hence they are useful for segmentation, registration,and identification of objects in a scene. The problem withedge extraction using gradient operators is that detection of edges is either in horizontal or in vertical directions, as thegradient operators take only the first order derivative of image.Shape feature extraction in image retrieval requires theextracted edges to be connected in order to reflect theboundaries of objects present in the image. Slope magnitudemethod[1] is used along with the gradient operators (Sobel,Prewitt, Robert and Canny)[1] to extract the shape features inform of connected boundaries. The process of applying theslope magnitude method is given as follows. First the imageneeds to be convolved with the Gx mask to get the x gradientand Gy mask to get the y gradient of the image. Then theindividual squares of both these gradients are taken. Square
60 http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 09, No.02, 2011
root of addition of the two squared terms gives the extractedconnected edges from the image as given in equation 1.
(1)
III.
BLOCK TRUNCATION CODING
Block truncation coding (BTC) is a simple image codingtechnique developed in the early years of digital imaging. BTChas played an important role in the history of digital imagecoding in the sense that many advanced coding techniqueshave been developed based on BTC or inspired by the successof BTC.This method first divides the image to be coded into smallnon-overlapping image blocks typically of size 4× 4 pixels toachieve reasonable quality. The small blocks are coded one ata time. For each block, the original pixels within the block arecoded using a binary bit-map the same Upper Mean Color(UM) size as the original blocks and two mean pixel values.The method first computes the mean pixel value of the wholeblock and then each pixel in that block is compared to theblock mean. If a pixel is greater than or equal to the block mean, the corresponding pixel position of the bitmap will havea value of 1 otherwise it will have a value of 0. Two meanpixel values one for the pixels greater than or equal to theblock mean and the other for the pixels smaller than the block mean are also calculated. At decoding stage, the small blocksare decoded one at a time. For each block, the pixel positionswhere the corresponding bitmap has a value of 1 is replacedby one mean pixel value and those pixel positions where thecorresponding bitmap has a value of 0 is replaced by anothermean pixel value.It was quite natural to extend BTC to multi - spectrumimages such as color images. Most color images are recordedin RGB space, which is perhaps the most well-known colorspace. As described previously, BTC divides the image to becoded into small blocks and code them one at a time. Forsingle bitmap BTC of color image, a single binary bitmap thesame size as the block is created and two colors are computedto approximate the pixels within the block. To create a binarybitmap in the RGB space, an inter band average image (IBAI)is first created and a single scalar value is found as thethreshold value. The bitmap is then created by comparing thepixels in the IBAI with the threshold value.
A.
Bit Calculation
Let X={R(
i,j),G(i,j),B(i,j)} where i=1,2,….m andj=1,2,….,n; be an m×n color image in RGB space.
Theinterband average image could be computed as IA={IB(i,j) }where i=1,2,---,m and j=1,2,-----,n and where
(2)The Threshold(T) is computed as the mean of IB(i,j).
(3)The Binary bitmap {BM(i,j)} with i=1,2,...,m and
j=1,2,…,n is computed as
(4)
B.
Upper mean and Lower mean calculation
After the creation of the bitmap, two representative (mean)colors are then computed. The two mean colors, Upper Meanand Lower Mean. The Upper Mean UM=(Rm1, Gm1, Bm1) iscomputed as following equations.
(5)
(6)
(7)The Lower Mean LM= (Rm2, Gm2, Bm2) is computed asfollowing equations:
(8)
(9)
(10)Now these Upper Mean and Lower Mean together will form afeature vector or signature of the image. For every imagestored in the database these feature vectors are computed andstored in feature vector table. Whenever a query image isgiven to CBIR, again the feature vector for query image willbe computed and then it will be matched with feature vectortable entries for best possible matches at given accuracy rate.Here we have used Direct Euclidean Distance as a similaritymeasure to compute the similarity measures of images forContent Based Image Retrieval applications.IV.
MULTILEVEL
BTCImage As seen above in section 2.4, the image data is dividedinto 6 parts using the 3 means calculated for each of the planes(R, G and B). This is called BTC - Level 1. Similarly, if theimage data is divided into 12 parts using the 6 means
61 http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 09, No.02, 2011
calculated of each of the 6 parts in Level 1, we obtain BTCLevel 2 data[21].Here the bitmap are prepared using upper and lower meanvalues of individual colour components. For Red colour
component, the bitmap “BMUR” and “BMLR” are generated
as given in equations 17 and 18. Similarly for Green colour
component “BMUG” & “BMLR” and for Blue colour components “BMUB” & “BMLB” can be generated.
(11)
(12)Using this bitmap the two mean colours per bitmap, one forthe pixels greater than or equal to the threshold and the otherfor the pixels smaller than the threshold are calculated. Theupper mean color UM (UUR, ULR, UUG, ULG, UUB, ULB)are given as follows.
(13)
(14)And the first two components of Lower Mean LM= (LUR,LLR, LUG, LLG, LUB, LLB) are computed using followingequations.
(15)
(16)These Upper Mean and Lower Mean together will form afeature vector for BTC
–
Level 2. For every image stored inthe database these feature vectors are computed and stored infeature vector table.Similarly the feature vector for BTC
–
Level 3 can be foundby extending the BTC
–
Level 2 till as shown in figure 20.Hence the image is divided into 24 parts using 12 meansgenerated from Level 2. Each plane will give the 6 elements of feature vector. For example for the Red plane we get ( UUUR,LUUR, ULUR, LLUR, UULR, LULR, ULLR, LLLR ).V.
PROPSED
CBIR
TECHNIQUES
The problem of having all the database images with samesize for image retrieval can be resolved using proposed Mask Shape BTC based CBIR methods. Here firstly, the shapefeatures of the image are extracted by applying slopemagnitude method on gradients of the image in vertical andhorizontal directions and then the BTC is applied on obtainedMask Shape images to have a shape feature vector withconstant size irrespective of size of the image considered.Even in Mask shape BTC based image retrieval four variationsare considered using different gradient operators.VI.
IMPLEMENTATION
The discussed image retrieval methods are implemented usingMATLAB 7.0 on Intel Core 2 Duo processor T8100(2.1 GHz)with 2 GB of RAM. To check the performance of proposedtechnique a database of 1000 variable sized images spreadacross 11 categories has been used[3]. Five queries wereselected from each category of images. Mean Squared Error(MSE) is used as similarity measure for comparing the queryimage with all the images in the image database. Let Vpi andVqi be the feature vectors of image P and Query image Qrespectively with size n, then the MSE can be given as shownin equation 17.
(17)To assess the retrieval effectiveness, we have used theprecision and recall as statistical comparison parameters forour proposed technique of CBIR. The standard definitions of these two measures are given by following equations.
retrieved imagesof number Total
retrieved imagesrelevant of Number
ecision
____
____
Pr
(18)
databaseinimagesrelevent of number Total
retrieved imagesrelevant of Number
call
______
____
Re
(19)VII.
RESULTS AND DISCUSSION
Figure 1: Crossover points for all levels of BTC for Canny Operator
Figure 1 shows a comparison between all the four levels of BTC by applying Canny operator. To get a betterunderstanding of the results figure 2 shows a zoomed versionof the same graph. From figure 2 we can see that level 3 givesthe best performance in comparison to the other levels. But wesee a drop in performance for level 4 due to the formation of null sets. Figure 3 shows a bar graph comparing the results of
0.20.250.30.350.40.45
Crossover point of Precision & Recall
Prewitt Robert Sobel
62 http://sites.google.com/site/ijcsis/ISSN 1947-5500
Search History:
Result 00 of 00
00 results for result for
p.
Improved Shape Content Based Image Retrieval Using Multilevel Block Truncation Coding
This paper presents improved content based image retrieval (CBIR) techniques based on multilevel Block Truncation Coding (BTC) using multiple threshold values. Block Truncation Coding based feature is one of the CBIR methods proposed using shape features…
(More)
511
Reads
0
Readcasts
0
Embed Views
More From This User
Notes
Load more
