IOSR Journal of Electroncs and Communcaton Engneerng (IOSR-JECE) e-issn: 78-834,p- ISSN: 78-8735.Volume 9, Issue, Ver. IV (Mar - Apr. 04), PP 0-07 Content Based Image Retreval Usng -D Dscrete Wavelet wth Texture Feature wth Dfferent Classfers Mohankumar C, Madhavan J (PG Scholar n Communcaton Systems, Adhyamaan College Of Engneerng, Hosur, Inda) (AP n Electroncs And Communcaton Engneerng, Adhyamaan College Of Engneerng, Hosur, Inda) Abstract:In ths paper we proposed dscrete wavelet transform wth texture for content based mage retreval. Our proposed method uses -D Dscrete Wavelet Transform for reducng the Dmensons of test mage and traned mages. Further gray level co-occurrence matrx s appled for all test and traned mages of LL components of level decomposed mages for extract the texture feature of the mages. Then smlar mages are retreved by usng dfferent dstance classfers. Expermental results are performed for Wang s database and t gves the mproved performance for homogenety wth energy property of GLCM of texture feature for Eucldean dstance method. Keywords CBIR, Texture, -D Dscrete Wavelet Transform, Eucldean dstance, Manhattan dstance. I. Introducton Content Based Image Retreval s to retreve an mage from the mage database when gven a query mage. Query Image s the users target mage for the searchng process. CBIR systems operate n two phases: ndexng and searchng. In the ndexng phase, each mage of the database s represented usng a set of mage attrbute, such as color, shape, texture and layout. Extracted features are stored n a feature database. In the searchng phase, when a user makes a query, a feature vector for the query s computed. Usng a smlarty crteron, ths vector s compared to the vectors n the feature database. The mages most smlar to the query are returned to the user. Rapd advances n hardware technology and growth of computer power make facltes for spread use of World Wde Web. Ths causes that dgtal lbrares manpulate huge amounts of mage data. Due to the lmtatons of space and tme, the mages are represented n compressed formats. Therefore, new waves of research efforts are drected to feature extracton n compressed doman. Wavelet transform can be used to characterze textures usng statstcal propertes of gray levels of the pxels comprsng a surface mage [4]. The wavelet transform s a tool that cuts up data or functons or operatons nto dfferent frequency components and then studes each component wth a resoluton matched to ts scale. In the proposed paper, we use -D Dscrete Wavelet transform wth textured feature mages. We also provde a retreval accuracy strategy for dfferent wavelets. The outlne of the paper s as follows. General structure of proposed system s revewed n secton. In secton 3 dscrete wavelet transform s descrbed. Secton 4 and 5 dscussed about Texture and dstance measures. Expermental results and conclusons are presented n secton 6 and 7 respectvely. II. General Structure Of Proposed CBIR System Fg. shows the basc block dagram of Content based mage retreval system. Our Proposed method compares the performance of Content based mage retreval usng DWT wth Texture for smlarty matchng of Eucldan dstance(l), Manhattan dstance(l) and Standard Eucldean dstance(std L) method. All tran mages decomposed usng dscrete wavelet transform. After DWT, we are takng only low frequency components (LL) of the mage for texture feature usng GLCM. Then the fnal feature vectors of all the tran mages are stored n the database. Same process s done for query mage. Fnally frst 5 smlar mages are retreved by usng L, L, std L method Query Image Image Decomposton Wth DWT GLCM for Texture Feature Smlarty Measures Matched Images Database Image Image Decomposton Wth DWT GLCM for Texture Feature Feature Database Fg. Block dagram of proposed system Page
III. Dscrete Wavelet Transform The Dscrete Wavelet Transform (DWT), whch s based on sub-band codng, s found to yeld a fast computaton of Wavelet Transform. It s easy to mplement and reduces the computaton tme and resources requred. The two-dmensonal DWT of an mage functon s(n,n ) of sze N x N may be expressed as N N W 0, k, k sn, n,,, 0 k k n n N N W n 0 n 0 N N, k, k sn, n n, n 0 N N n 0 n 0 0, k, k Where = {H, V, D} ndcate the drecton ndex of the wavelet functon. As n one-dmensonal case. 0 represents any startng scale, whch may be treated as 0 =0. Gven the above two equatons are two-dmensonal DWT. The specalty of DWT [6] comparng to other transforms s tme and frequency characterstcs whch applcaton of t wldly n many dfferent felds. The flow chart of the Dscrete wavelet transform sub band codng on the dgtal mage s shown n fgure, here L refers to low frequency component, H refers to hgh frequency and the number and refer to the decomposton level of the Dscrete wavelet transform. The result of the -D Dscrete Wavelet Transform from level one to level three s shown n fgure 3. The sub mage LL s the low frequency component, t s the approxmate sub mage of the orgnal mage; the sub mage HL s the component of the low frequency n horzontal drecton and the hgh frequency n vertcal drecton, t manfests the horzontal edge of the orgnal mage; the sub mage LH s the component of the hgh frequency n horzontal drecton and the low frequency n vertcal drecton, t manfests the vertcal edge of the orgnal mage; the sub mage HH s the hgh frequency component, t manfests the oblque edge of the orgnal mage. It s shown that most energy of the orgnal mage s contaned n the LL low frequency regon. And the other regon n the same sze reflect edge feature of the mage n dfferent angles. Here we use the -D Dscrete Haar wavelet transform for the decomposton of the mages. LL LH LL LH LH Input Image HL HH HL HH HL HH Fg. Flow chart of the DWT sub band codng 50 0 40 00 60 50 80 00 00 50 50 00 50 00 50 300 350 0 0 40 60 80 00 0 40 60 80 (a) (b) Fg. 3 (a) and (b) shows decomposton of mages by level and. Fg. 3(a) and (b) shows the level one and level two decomposton of the mage. IV. Texture Feature Extracton An mportant feature of an mage s texture. To descrbe the texture of the regon three approaches are used n mage processng these are statstcal, structural and spectral. Statstcal approaches specfy the characterzaton of the textures by smooth, coarse, grany, slky and so on. The common second order statstc s gray level co occurrence matrx. Page
4.. Gray Level Co-Occurrence Matrx (GLCM) Gray Level Co-occurrence Matrx (GLCM) method s based on the condtonal probablty densty functon. GLCM ntroduced by Haralck. It contans the nformaton about the postons of pxels that havng smlar gray level values. Co-occurrence matrx functon represents the drecton and dstance. In the gven drecton and dstance we can calculate the symbolc gray level pxel,. That can be expressed as the number of co occurrence matrx p, d,. Element p, d, p A GLCM s represented, as a matrx. In whch the number of rows and columns s equal to the number of gray levels n the mage. The matrx element P(, d, θ) s the relatve frequency wth whch two pxels, separated by dstance d. The drecton specfed by the partcular angle (θ). Texture feature are computed from the statstcal dstrbuton. Ths can be observed combnaton of ntensty at specfed poston relatve to each other n the mage. Here texture feature are extracted by usng gray level co-occurrences matrxes of these three propertes. 4.. Homogenety 4.. Energy 4..3 Correlaton 4..Homogenety It returns a value that measures the closeness of the dstrbuton of elements n the GLCM. Its range s [0 ]. Homogenety s for a dagonal GLCM. H N, 0, d p, P,, d 4.. Energy The sum of squared elements n the gray level co-occurrence matrx called Energy. Its range s [0 ].Energy s for a constant mage. E N P,, 0 4..3 Correlaton Returns a measure of how correlated a pxel s to ts neghbor over the whole mage. Its range s [- ]. Correlaton s or - for a perfectly postvely or negatvely correlated mage. Correlaton s NaN for a constant mage. C,, p, V. Dstance Measures Dstance measures such as the Eucldean, Manhattan and Standard Eucldean dstance have been used to determne the smlarty of feature vectors. In ths CBIR system Eucldean dstance, Standard Eucldean dstance and also Manhattan dstance s used to commonly to compare the smlarty between the mages. Dstance between two mages s used to fnd the smlartes between query mage and the mages n the database. 5. Eucldean dstance Eucldean dstance square root of the sum of the squares of the dstance between correspondng values. d 5. Manhattan dstance It computes the dstance that would be travelled to get from one data pont to other. The sum of the dfference of ther correspondng samples. n d n x y x y 3 Page
5.3 Standard Eucldean Dstance Standardzed Eucldean dstance means Eucldean dstance s calculated on standardzed data. Standardzed value = (Orgnal value - mean)/standard Devaton d n s x y VI. Smulaton Result And Dscusson Smulaton results are performed wth MATLAB. Ths database conssts of 000 mages each classes have 00 mages for the expermental purpose we took ten mages from each class totally 00 mages. Ths experment gves the performance comparson result of content based mage retreval for the metrc L, L and StdL.Here feature extracton s done by Haar wavelet of level wth texture feature. Here we combne the GLCM propertes of homogenety wth energy and energy wth correlaton to extract texture feature the comparson are done for ths two combned propertes wth dfferent dstance classfers. Fg. 4 shows the Query mage and retreved mages by smlarty measures of the Eucldean dstance for homogenety and energy property. Fg. 4 Smulaton Result for -D DWT wth Texture for Eucldean Dstance Fg. 5 shows the Query mage and retreved mages by smlarty measures of the Manhattan dstance for homogenety and energy property. Fg.5 Smulaton Result for DWT wth Texture for Manhattan Dstance 4 Page
Fg. 6 shows the Query mage and retreved mages by smlarty measures of the Standard Eucldean dstance for homogenety and energy property. Fg.6 Smulaton Result for DWT wth Texture for Standard Eucldean Dstance Fg.7 shows the Query mage and retreved mages by smlarty measures of the Eucldean dstance for correlaton and energy property. Fg. 7 Smulaton Result for -D DWT wth Texture for Eucldean Dstance Fg. 8 shows the Query mage and retreved mages by smlarty measures of the Manhattan dstance for correlaton and energy property. Fg. 8 Smulaton Result for -D DWT wth Texture for Manhattan Dstance 5 Page
Fg. 9 shows the Query mage and retreved mages by smlarty measures of the Standard Eucldean dstance for correlaton and energy property. Fg. 9 Smulaton Result for -D DWT wth Texture for Standard Eucldean Dstance Retreval Accuracy s calculated by No of Relevant Images Retreved Accuracy Total No of Images Retreved Ths table shows the retreval accuracy for each class of the CBIR system. Class Name Manhattan dstance L TABLE. RETRIEVAL ACCURACY Energy + Correlaton Homogenety + Energy Standard Eucldean Manhattan Eucldean Eucldean Dstance dstance Dstance Dstance L L L Std L Standard Eucldean Dstance Std L People 70 76 70 7 76 74 Beach 78 8 78 9 9 8 Buldng 8 8 78 9 9 9 Buses 94 96 9 94 94 84 Dnosaur 96 96 94 00 00 00 Elephant 88 90 84 98 98 9 Horses 98 96 96 98 98 98 Roses 88 86 84 76 80 86 Mountan 86 86 80 98 96 84 Food 88 88 8 78 80 84 Average Retreval Accuracy 86.8 87.8 83.8 89.8 90.6 87.6 VIII. Concluson It s observed that proposed method mproves the Content Based Image Retreval wth -D Dscrete Haar Wavelet wth gray level co-occurrence matrx whch s used to extract the texture feature of the mages. Here we compared the result wth dfferent dstance measure classfers and also wth dfferent GLCM property. Table shows performance comparson of Homogenety wth Energy and Energy wth Correlaton wth dfferent dstance classfer n that one Eucldean dstance gves best performance than Manhattan and Standard Eucldean dstance for both Homogenety wth Energy and Energy wth Correlaton. When comparng wth combned property of homogenety wth energy and energy wth correlaton homogenety wth energy gves the better performance. References [] PotrPorwk, AgneszkaLsowska, The Haar Wavelet Transform n Dgtal Image Processng: Its Status and Achevements. [] Phang Chang, PhangPau, Smple Procedure for the Desgnaton of Haar Wavelet Matrces for Dfferental Equatons, Proceedng of the Internatonal Multconference of Engneers and Computer Scentsts 008 Vol II, March 008. [3] Patrck J. Van Fleet, Dscrete Haar Wavelet Transform, PREP, Wavelet Workshop 006. [4] N. GnaneshwaraRao, Dr. V. Vaya Kumara, V Venkata Krshna, Texture Based Image Indexng and Retreval, IJCSNS Internatonal Journal of Computer Scence and Network Securty, VOL.9 No.5, May 009 [5] HossenNezamabad-pour and SaedSaryazd, Obect Based Image Indexng and Retreval n DCT Doman usng Clusterng Technques, World Academy of Scence, Engneerng and Technology 3 005 6 Page
[6] Hong Wang, Su Yang, We Lao, An Improved PCA Face Recognton Algorthm Based on the Dscrete Wavelet Transform and the Support Vector Machnes, Internatonal Conference on Computatonal Intellgence and Securty Workshops 308-3,007. [7] http://www.fp.ucalgary.ca/mhallbey/tutoral.html [8] Prof. SomnathSengupta, Mult Resoluton Analyss Verson ECE IIT, Kharagpur [9] J.Madhavan, K.Porkumaran, Performance Comparson Of PCA, DWT-PCA And LWT-PCA For Face Image Retreval, Computer Scence & Engneerng: An Internatonal Journal (CSEIJ), Vol., No.6, December 0 7 Page