ISS: 2278 323 Internatonal Journal of Advanced Research n Computer Engneerng & Technology (IJARCET) Local Tr-drectonal Weber Rhombus Co-occurrence Pattern: A ew Texture Descrptor for Brodatz Texture Image Retreval Venkata Satya Kumar Gangavarapu, Gopala Krshna Mohan Pllutla Abstract A new feature extracton method called Local Tr-drectonal Weber Rhombus Co-occurrence Pattern (LTrWRCoP) for texture mage retreval s presented n the paper. Most of the local bnary pattern (LBP) varants extract the local nformaton based on dfference of current pxel wth ts neghborhood pxels but they gnore the orgnal ntensty of the stmulus. The proposed LTrWRCoP not only explores the nter relatonshp among the neghborhood pxels but also consders the orgnal ntensty of stmulus for extractng the local nformaton structure. Further, gray level co occurrence matrx (GLCM) s used to get the co occurrences of pxel pars n local pattern map as t s more robust than the frequency of patterns obtaned usng hstogram. The proposed method also examne the co-occurrence of pxel pars n varous drectons and dstances. The expermental results on the Brodatz texture database reveals the superorty of the proposed method to the other methods n terms of average precson and recall rates Index Terms GLCM, Image retreval, Local bnary pattern, Pattern recognton,texture I. ITRODUCTIO Enhancements n multmeda technology lead to exponental growth n the sze of mage repostores. Managng and archvng these databases became a herculean task. Texture s an mportant low level feature of an mage. Repettve blocks or a smlar pattern n an mage ndcates the presence of texture. Local bnary patterns (LBP) proposed by Ojala et al [] showed promsng results n texture feature extracton and object trackng. For texture feature extracton several methods have been used ncludng local bnary patterns, local ternary patterns, dscrete wavelet transform, Gabor flters etc. The concept of gray level co occurrence matrx (GLCM) s ntroduced by Haralck et.al to extract statstcal features for texture mage classfcaton [2]. The lterature ponts the fact that the GLCM provdes the spatal co-relaton of pxels n the mage and t s useful n texture feature extracton. The GLCM s used for better feature descrpton n texture descrptors. Further, local tetra patterns, local ternary co-occurrence patterns and modfed colour motf co-occurrence matrx for mage ndexng and retreval are proposed by Mural et al [3-8]. center symmetrc local bnary co-occurrence pattern for bomedcal and texture mages usng GLCM s proposed by Mansha et al. A local tr-drectonal pattern for mage retreval s presented n [9]. The combnaton of hstogram of colour and local Rhombus pattern for object trackng s presented n [0]. Most of the pattern based technques lke LBP, Local ternary patterns (LTP), Center symmetrc local bnary patterns (CSLBP) encodes the dfference of pxels to obtan local nformaton structure but gnores the orgnal stmulus ntensty[-6]. Weber local descrptor proposed by Je Chen et al accounts for the orgnal ntensty of the stmulus resemblng to human percepton and uses the hstogram to extract the frequency nformaton of local pattern map[7]. The center symmetrc local bnary co-occurrence pattern(cslbcop) proposed by Mansha et al encodes the dfference of pxel n symmetrc neghbourhood for local pattern map and obtans the co-occurrence pxel par n local pattern map for better feature extracton but omts the orgnal ntensty stmulus. The proposed Local tr-drectonal Weber Rhombus co-occurrence pattern not only consders the orgnal ntensty stmulus but also explots the co-occurrences of pxel pars n local pattern map wth GLCM. The experments conducted on Brodatz texture mage database shows the effcency of proposed method compared CSLBP, CSLBCoP n terms of average precson rate and average recall rate. A. Weber Local Descrptor II. RELATED WORK Jn Chen et al proposed weber local descrptor as functon of dfferental exctaton and gradent orentaton. The dfferental exctaton ( ) of a pxel wth eght neghbourhoods (p=8) s defned as n [7] p ( y yc) ( yc ) arctan yc y s neghbourhood pxel and y c s center pxel Gradent orentaton (G) s defned as G ( yc) medan( G), where p 0,, 2,... () (2) 2 All Rghts Reserved 206 IJARCET 75
www.jarcet.org 752 ISS: 2278 323 Internatonal Journal of Advanced Research n Computer Engneerng & Technology (IJARCET) dfference G represents the angle of a gradent Fg (a): Example sub mage (b): Local Rhombus pattern example wndow yr( 4) y G arctan (3) yr( 6) y R( 2) p y,( 0,, 2,... ) are neghbours of a current 2 pxel, R( y) s obtaned usng modulus operaton.e., R( y) mod( y, p) where p s the number of neghbours. The Weber local descrptor s gven by WLD(, G). B. Local Bnary Patterns The local bnary pattern for p neghbourhood and d radus s defned as n Ojala et al p n p, d n c n0 LBP 2 S( y y ) (4) 0, x 0 Sx ( ), x 0 (5) y c y -center pxel, m -neghbourhood pxel ntenstes. Hstogram of LBP map s obtaned usng the equaton as m n H( L) S pattern( y, y ), L (6) pattern, j S(, j) 0, else (7) 2 xx2 III. PROPOSED METHOD, p L 0,(2 ) In WLD, hstogram that express the frequency of each pattern s consdered but mutual occurrences of pattern s gnored. as, CSLBCoP consders the mutual occurrences of pattern and gnores the orgnal stmulus of ntensty. The proposed method combnes the best of both the methods by consderng the mutual occurrence of pattern to represent the mage features and the orgnal stmulus ntensty. Consder a sub mage havng center pxel y c wth p=8 neghbourhood pxels as shown n the Fg. Fg(c)-(f) shows the consderaton of tr-drectonal pxels n a Rhombus pattern for a gven sample wndow Frst, calculate the dfferental exctaton of pxels y, y 3, y 4, and y 5 usng the formulae gven below ( y y ) ( y y ) ( yc y ) d, 3,5,7 y y y (8) ( y ) arctan( d ) (0) ( y ) s the dfferental exctaton of th neghbourhood pxel. If ( y ) 0 means that the current pxel y s brghter than the surroundng pxels n the gven drecton otherwse y s lghter than the surroundng pxel. The pattern map LTrWRP s defned for a gven 3x3 wndow as (), ( y ) 0 f( ( y )) 0, ( y ) 0 (9) 3 LTrW( y ) ( y ), ( y ), ( y ), ( y ) (2) C ( y y ) ( y 7 y ) ( yc y ), d y y y 3 5 7 0 2 f ( y ) 2 f ( y ) LTrWRP( yc ) (3) 2 3 2 f ( y5) 2 f ( y7) The Patterns obtaned from LTrWRP, ranged from 0 to5. The tr-drectonal weber rhombus pattern s obtaned for the gven nput mage. Eght neghbourhood pxels wth a radus r= are consdered for the pattern. After pattern map, the range of ntensty n the pattern vares from 0 to 5. Gray level co-occurrence matrx s used to obtan the occurrence of pxels pars n the LTrWRP pattern mapped mage. The GLCM can be extracted n varous drectons and dstances for the gven mage. In the proposed method, four combnatons of GLCM are demonstrated as follows Combnaton: Four GLCM of dstance wth angles 0 0,45 0,90 0 and 35 0 are extracted. Combnaton2: Four GLCM of dstance 2 wth angles 0 0,45 0,90 0 and 35 0 are extracted. Combnaton3: Two GLCM of dstance wth angles 0 0,45 0 and two GLCM of dstance 2 wth angles 0 0,45 0 Combnaton4: Two GLCM of dstance wth angles 0 0,90 0 and two GLCM of dstance 2 wth angles 0 0,90 0 As the ntensty values vary from 0 to 5(Total 6 ntenstes) n the pattern map, the length of GLCM matrx s
ISS: 2278 323 Internatonal Journal of Advanced Research n Computer Engneerng & Technology (IJARCET) 6 x 6 and each combnaton has four such GLCMS. Therefore, the feature vector length wll be 4x6x6=024. IV. PROPOSED SYSTEM FRAMEWORK A. Feature extracton The algorthm to extract features from a gven mage as follows Input: Image Output: Feature vector Step: Input the gray scale mage or convert the mage to nto gray scale f t s RGB mage Step2: Apply Local tr-drectonal Weber Rhombus Pattern to get the pattern of the gven mage. Step3: Apply any one combnaton of GLCM as explaned n the prevous secton at varous dstances and angles. Step4: convert the four 6x6 matrces nto vectors obtaned from prevous step Step5: concatenate all four vectors obtaned n step4 nto a sngle vector to form the feature vector. B. Smlarty measure The query mage feature vector s represented f ( f, f, f,... f )., L s the length of the by Q 2 3 L feature vector obtaned after feature extracton. The features vectors n the database are represented f ( f, f,... f ), represent the number of by DB DB DB 2 DB mages n the data base. The goal of smlarty measure s to retreve n top matches for the gven query mage from the feature database by measurng the dstance between query mage features and mage features n the database. For smlarty measure, d dstance metrc s used and t s computed as follows d( Q, DB) (4) L f DB f f Q f DB Q fdb feature vector of th mage n the database s fq s feature vector of query mage. d( Q, DB) - dstance functon. V. EXPERIMETAL RESULTS The performance of the proposed method s compared to the exstng methods n terms of average precson rate (APR) and average recall rate (ARR).The formulae for precson and recall as follows The precson and recall for th mage n the database s gven by P ( ) and R ( ) respectvely wth number of mages retreved for each query mage. R P ( ) (5) T R s number of relevant mages retreved, T s total number of mages retreved R R ( ) (6) R s number of relevant mages retreved, D s total number of mages n the database. The average precson ( ) D AP and average th recall AR for j category wth number of mages are determned by usng the formula. AP( j) P( ) (7) AR( j) R( ) (8) Average precson rate (APR) and average recall rate (ARR) for a gven database wth categores are obtaned by usng the formulae. 2 APR AP() (9) A. Experment# 2 2 ARR AR() (20) 2 Brodatz texture database [8] s consdered for the experment. It conssts of 2 mages of sze 640 640. For the experment, each mage s dvded nto 25 sub mages of sze 28 28. Therefore 2 25=2800 mages are consdered for evaluaton of performance. Each mage n the database s gven as query mage. The average precson, average recall s computed for dfferent methods usng combnaton (GLCM wth d=, angles 0 0,45 0,90 0 and 35 0 ) and graphs are plotted as shown n Fg2. From Fg2 t s evdent that the proposed method (LTrWRCoP) outperformed the other exstng technques n terms of ARR, ARP. From Fg3 t s evdent that combnaton4.e. Two GLCM of dstance wth angles 0 0,90 0 and two GLCM of dstance 2 wth angles 0 0,90 0 resulted n mproved recall rates compared other combnatons. as, the change n precson s small compared to recall rates for varous combnatons All Rghts Reserved 206 IJARCET 753
www.jarcet.org 754 ISS: 2278 323 Internatonal Journal of Advanced Research n Computer Engneerng & Technology (IJARCET) Fg2: Average recall percentage, average precson percentage curves for Brodatz texture database Fg3. LTrWRCoP for Brodatz texture database for varous combnatons of GLCM wth dfferent drectons and dstances VI.COCLUSIO A novel mage retreval algorthm called Local Tr-drectonal Weber Rhombus Co-occurrence Pattern s proposed n ths paper. Each pxel s compared to ts most adjacent neghbourhood and center pxel n a three drectons for local nformaton extracton. The magntude of dfferental exctaton s consdered to obtan salent features wthn a local neghbourhood to smulate human beng s percepton of patterns. Gray level co-occurrence matrx s used to explore the mutual co-occurrence of pattern pars, whch s robust compared to the hstogram technque. Expermental results conducted on Brodatz texture database ndcate that the proposed method s superor to the exstng patterns LBP, CSLBP and CSLBCoP n terms of average precson rate and average recall rate. REFERECES [] S.A.Orjuela Vargaa,J.P.Yanez Puentes, and W. Phlps, Local bnary patterns: ew varants and applcatons, vol.506, pp.85-2, DOI: 0.007/978-3-642-39289-4_4. [2] R.M.Haralck,K.Shanmugam,I.Dnsten, Textural features for mage classfcaton,ieee Trans.Sys.Man.Cyber,vol.6,pp.60-62,973 [3] M.Subrahmanyam,R.P.Maheswar,andR.Balasubramanan, Drectonal bnary wavelet patterns for bomedcal mage ndexng and retreval, Journal of Med.syst, vol.36,pp.2865-2879,202. [4] M.Subrahmanyam,R.P.Maheswar,andR.Balasubramanan, Local maxmum edge bnary patterns: a new descrptor for mage retreval and obejcet trackng, Sgnal processng(elsever),vol.92,pp.467-479,202. [5] M.Subrahmanyam,R.P.Maheswar,andR.Balasubramanan, Local tetra patterns: a new feature descrptor for content based mage retreval, IEEE Trans.Image Process.,vol.2,pp.2874-2886,202. [6] M.Subrahmanyam,Q.J.Wu, Sphercal symmetrc 3D local ternary patterns for natural,texture and bomedcal mage ndexng and retreval, Journal of euro computng, vol.49,pp.502-54,205. [7] M.Subrahmanyam,Q.J.Wu, Local Ternary Co-occurrence Patterns: A ew Feature Descrptor for MRI and CT Image Retreval, Journal of euro computng(elsever),vol.9,pp.399-42,203.
ISS: 2278 323 Internatonal Journal of Advanced Research n Computer Engneerng & Technology (IJARCET) [8] Subrahmanyam. Murala, QM.J. Wu, R.P. Maheshwar, and R. Balasubramanan, Modfed color motf co-occurrence matrx for mage ndexng and retreval, Comput. Electr.Eng,vol.39,pp.762-774,203. [9] M. Verma, B. Raman, Center symmetrc local bnary co-occurrence pattern for texture, face and bo-medcal mage retreval, J. Vs. Commun. Image Represent,vol.32,pp.224-236,205. [0] Mansha Verma, Balasubramanan Raman,"Local tr-drectonal patterns: A new texture feature descrptor for mage retreval, Dgtal sgnal processng, vol.5, pp. 62 72, Aprl 206 [] Mansha Verma, Balasubramanan Raman, object trackng usng jont hstogram of colour and local Rhombus pattern, IEEE Internatonal Conference on Sgnal and Image Processng Applcatons (ICSIPA),205. [2] A.W.Smeulders,M.Worrng,S.Santn, A.Gupta,and R.Jan, Content based mage retreval at the end of the early years, IEEE Ttrans.Pattern.Anal.Mach.Intell.,vol.223,pp.349-380,2000. [3] T.Ojala,M.Petkanen,and D.Harwood, A comparatve study of texture measures wth classfcaton based on featured dstrbutons, Pattern Recognton, vol.29,pp.5-59,996 [4] T.Ojala,M.Petkanen,and T.Maenpaa, Multresoluton gray-scale and rotaton nvarant texture classfcaton wth local bnary patterns, IEEE Trans.Pattern Anal.Mach.Intell.,vol.24,pp.97-987,2002. [5] Z.Guo,L.Zhang, and D.Zhang, Rotaton nvarant texture classfcaton usng LBP varance wth global matchng, Pattern Recognton, vol.43,pp.706-79,200. [6] X.Qan,X.S.Hua,P.Chen,and L.Ke, PLBP: an effectve local bnary patterns texture descrptor wth pyramd representaton, Pattern Recongnton, vol.44,pp.2502-255,20. [7] Je Chen, Shngan Shan,Guoyng Zhao, and Xln Chen, A Robust Desrptor based on Weber s Law, IEEE Transactons on pattern analyss and machne ntellgence, vol.32,pp.705-720,2009 [8] Brodatz texture database, [onlne]http://multbandtexture.recherche.usherbrooke.ca/orgnal_broda tz.html. Mr. G V SATYA KUMAR obtaned hs B.Tech degree from JT Unversty, Hyderabad n year 2002 and M.Tech degree from AU, Guntur n the year 2008. Presently he s pursung Ph. D n mage processng under the gudance of Dr P.G. Krshna Mohan. Hs areas of nterests are mage retreval, object trackng. Dr. P. G. Krshna Mohan presently workng as Professor n Insttute of Aeronautcal Engneerng, Hyderabad. He Worked as Head of ECE Dept., Member of BOS for ECE faculty at Unversty Level, Charman of BOS of EIE group at Unversty level, Charman of BOS of ECE faculty for JTUCEH, Member of selecton commttees for Kaktya, agarjuna Unversty, DRDL and convener for Unverste a Hdan commttees. He has more than 43 papers n varous Internatonal and atonal Journals and Conferences. Hs areas of specalzaton are Sgnal Processng, Sgnal Estmaton, Probablty Random Varables and Communcatons. All Rghts Reserved 206 IJARCET 755