Road Surface Crack Identification by Using Different Classifiers on Digital Images

Transcription

1 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) Road Surface Crack Idetificatio by Usig Differet Classifiers o Digital Iages Dr YDAR TOOSSIAN SANDIZ, OSIN GASMZAD TRANI, ADI ADIZAD Sarood Uiersity of Tecology, lectrical, Ciil gieerig Faculty 7 T Tir Square, P.o.Box , Sarood, IRAN Abstract: I tis paper differet classifier are used to idetifyig differet type of cracks o road surface. As our experiece sows Regio Growig Classifier (RGC) etod ca be used to diide all surface road iages i two ai groups. First group coers alligator ad block cracks. Logitudial, traserse cracks ad oter kid of distress are put i secod group. I first group, waelet Statistic Feature Classifier (WSFC), ertical ad orizotal istogra ad proxiity are used for classificatio. Tey elp to judge about te kid of crack based o digital iage fro road surface. istogra, RGC ad proxiity are classifiers wic are used i secod group. Multi layer Perceptro eural etwork is used to judge about te cracks. Keywords: Road Crack, Regio Growig Classifier (RGC), Statistic Feature Classifier (WSFC), Multi Layer Perceptro, Patter recogitio Itroductio igways ad roads are a ajor public asset i all coutries. To efficietly aage tese assets road autorities eed accurate, up-to-date iforatio o te coditio of teir igway ad road etworks. For exaple te aiteace ad reabilitatio of igway paeets i te uited state requires oer 7 billio dollars a year. Coetioal isual ad aual paeet crackig aalysis etods are ery costly, tie cosuig, dagerous, labor itesie ad subjectie. Autoatic oitorig of soe aspects of road coditio, for exaple rougess ad skid resistace, as bee carried out for a uber of years. oweer, oe of te ost iportat road quality idicators, te extet ad type of crackig, as up util ow bee easured oly by isual ispectio. Te result is tat oly ery sparse saplig as bee carried out, at a ery ig cost per kiloeter, ad ery little iforatio as terefore bee aailable i tis iportat aspect of road coditio. Te ai idea of digital iage processig etods is based o te fact tat te crack pixels i paeet iages are darker ta te surroudigs ad cotiuous [], [2]. Based o researcers work te distress o paeet ca be categorized as follow [3-5]: A logitudial crack wic is appears alog te igway. Traserses crack is a crack perpedicular to te paeet ceterlie. Alligator crack wic is a series of itercoected cracks wit ay sides ad sarp agled pieces. Block crack as a patter of rectagular pieces of road surface fro traserse cracks.

2 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) Oter distress, suc as a-oles, costructio plates, etc. Tis work itself ca be categorized as digital iage patter recogitio. Tere is a wide rage of patter recogitio approaces. Tey are categorized i two ai groups as statistical ad structural etods. I statistical etods, te iage is processed as a wole ad classified based o te distributed of te black pixels. I structural etod, te iage is expressed as copositios of structural uits. Te patter is recogized by atcig its structural represetatio wit tat of a refereces patter [6]. I te first step te crack is classified as wole ad i te secod step judge is based o structural etod. I te followig sectios ad subsectios te teory of our etod ad experietal results are discussed. Fig. Flowcart of proposed etod 2 Metod flowcarts Fig. sows te flowcart of proposed etod. ac block is explaied briefly. (a) (b) 2. Iage Acquisitio A digital ady ca is used to take road iage. Te degree of ady ca wit orizotal axes is 30 degree ad it is.5 eter aboe te surface. Te fil is fed to te lap-top ia i-lik. A coputer progra diided te fil i separate iages. Four kid cracks iages are sow i fig 2. Te origial iage as 640 by 480 pixels. ac pixel i orizotal directio represeted 0.3 c ad i ertical 0.65 c log. I oter word eac origial iage coers a area wit 46 by 44 c of te road. Our experiece sows te best iages are take i suy weater two or tree ours after rai. As te gray leel of te cracks is uc differet fro te backgroud of surface, suc iages are uc better ta usual iages. (c) 2.2 Preprocessig We are iterested to put all pixels i te iages i two groups, backgroud ad cracks pixels. It eas te backgroud ad oisy pixels ae to be sow i oe group. To aciee tis goal, a coputer progra first caged eac iage to gray scale te by usig proper tresold all iages caged to two leel as zero ad 255 gray scale (biary iage). Te produced biary iage is fed to ext block. (d) Fig.2 Scaled crack iages (a) Alligator, (b) Blocks, (c) Logitudial, (d) Traserse

3 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) 2.3 Regio Growig Classifier (RGC) I our approac, RGC plays a iportat ad critical role. Te pre-processed iage is classified as class A ad class B. Te adatage of suc a procedure is breakig te proble ito a uber of sipler oes. Te basic idea beid te RGC etod is to seget uderlyig iage ito soe disjoit regios ad cout te uber of regios wic are produced as te desirable output. Te process of forig a regio-based iage descriptio (or approxiatio) is referred to as segetatio. Iage segetatio ca be described as te partitioig of a iage ito a uber of disjoit segets or regios based o pixel grey leel caracteristics. Tese regios ay be sall eigboroods or ee sigle pixels. Tere are tree ai classes of iage segetatio teciques: statistical classificatio, edge detectio ad regio growig. Te segetatio to be used ere is of regio growig type. A regio is defied as a area i te iage wose pixels sare coo properties suc as siilar grey leel alues wic eclosed i a closed cotour. I oter word Regio growig is te process of joiig eigborig pixels ito larger regios based o tese properties [7]. 2.4 Waelet Statistic Feature (WSF) Classifier Waelet trasfor is a powerful ad faous ulti-resolutio aalysis wic as receied a lot of attetio at recet years. It offers a extra adatage, wic i soe cases ca be beeficially exploited. Its ultiresolutio properties cofor to te way perceptio is acieed by uas, troug teir earig ad isual systes. By usig tis trasfor at discrete case we ca aalysis ad iterpret te iput iage at ulti scales ad directios ad exploit te adatages of its ulti-resolutio properties [8]. I our proposed etod, waelet trasfor plays te key role i class A. It is resposible for puttig te iput iage ito oe of te logitudial or traserse crack type. For atteptig to tis goal, we are used oe of te ost popular statistical features of te waelet trasfor aely, eergy of eac decoposed iage wic is described latter. Let I ( x, y ) idicate te gray leel of iput iage. Oe leel of waelet decopositio o I ( x, y ) results i four sub iages: first soot sub iages I LL ( x, y ) wic represets te coarse approxiatio of te iput iage, ad tree detail sub iages I L ( x, y ), I L ( x, y ) ad I ( x, y ) wic represet te orizotal, ertical, ad diagoal directios of te iage, respectiely. Furter, let I LL ( x, y ) represets te soot sub iage at resolutio leel ad 0 I LL I( x, wic is te origial iage. Te te decopositio of I LL ( x, y ) results i four sub iages I + LL, I + L, I + L ad I + at resolutio leel + eac of + size Fig. 3 sows a tree leel decopositio of a saple logitudial crack. Fig. 3 Tree decopositio leels wit discrete waelet trasfor Now, we itroduce te ergy Statistic Feature (SF) of waelet trasfor ad use it i our WSF etod. Te eergy of eac decoposed sub iage is calculated as follows:

4 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) Te eergy of te soot sub iage wic is i fact a coarse approxiatio to its origial iage at leel is gie by: s x y ( ) LL 2 [ I ] () Te eergy of te orizotal detail sub iage at leel is: ( ) 2 (2) [ I ],,2, 3 x y L Te eergy of te ertical detail sub iage at leel is: ( ) 2 (3) [ I ],,2, 3 x y L Te eergy of te diagoal detail sub iage at leel is: ( ) 2 (4) [ I ],,2, 3 d x y Ad te oralized eergy of eac decoposed sub iage is defied as: s d (5) (6) (7) Now we defie as te ratio of oralized ertical eergy to te oralized orizotal eergy at sub bad as follows: (8) It is obious tat for a logitudial crack, bigger ta ad for a traserse crack is bigger ta. So if is bigger ta, te WSF suggests tat te iput iage is a logitudial crack ad if it is less ta, te WSF suggests tat te iput iage is a traserse crack. 2.5 istogra Fro te pre-processed iput iage two kids of istogras are calculated, a ertical istogra ad a orizotal istogra. Te ertical istogra of a biary iage is defied as te uber of o-zero alues i eac colu ad te orizotal istogra of a biary iage is defied as te uber of o-zero alues i eac row. Based o tese defiitios for a gie iage forula (9) ad (0) are itroduced as ertical istogra ad orizotal istogra, respectiely. I( i, j) i j, K, ( i) I( i, j) j i, K, (9) (0) istogras sow a clear patter of a crack. If a crack is deeloped i a logitudial directio, tere is a clear peak i te ertical istogra ad as a soot or costat ariatio i te orizotal istogra. Istead, if a crack is deeloped i a traserse directio, tere is a clear peak i te orizotal istogra ad as a costat ariatio i te ertical istogra. If a crack is a alligator crack, te peaks ca be foud i bot ertical ad orizotal directios. For a block crack, peaks ca be also foud i bot istogras but wit a agitude lower ta tose of a alligator crack. To fid a way for usig te abilities of istogra etod we proposed usig te ea alue of eac ertical istogra ad orizotal istogra as follows: is

5 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) µ µ () (2) Were µ ad µ are te ea or dc alue of ertical istogras ad orizotal istogras, respectiely. Oe adatage of usig te ea alues of istogra is tat it gies a positio iariat easure of te iput iage because if te crack is sifted across te orizotal or ertical directio te µ ad µ are ot caged. 2.6 Proxiity Te ea alues of orizotal istogra for logitudial ad traserse cracks are ery close to eac oter ad so te segregatio betwee tese two cracks is poor. To reedy tis proble we defie a proxiity easure as a alteratie as ertical Proxiity ad orizotal Proxiity by equatio 3 ad 4, respectiely. j i ( j + ) ( i + ) ( i) (3) (4) Fro te aboe equatios it ca be see tat proxiity is coputed by accuulatig te differeces betwee adjacet istogra alues. Te low alue of proxiity idicates tat tere is little differece betwee ay of te colus or rows for te iput iage. It is clear tat te ertical proxiity for a logitudial is bigger ta ertical proxiity i a traserse crack ad ice ersa. 2.7 udge Now, we are ready to produce a suitable feature ector for eac described class. A Multi Layer Perceptro (MLP) is used to classify eac iage. Te feature ector i class A is F A () (2) (3) {,,,, µ, } µ, (5) ad i Class B is {,, } F µ µ, (6) B, 3 xperietal results As explaied i sectio a alligator crack is a series of itercoected cracks, wic as ay sided ad sarp-agled pieces wereas a block crack is a patter of rectagular pieces of aspalt surface. So we expect tat te coplexity ad te uber of regios i a alligator crack are bigger ta a block crack. It is clear tat te output of te RGC or for alligator ad block cracks is ore ta tree oter kids. To sow our clai, table represets te aerage output of RGC for about 45 kids of alligator cracks, 45 kids of block cracks, 35 kids of logitudial cracks, 30 kids of traserse cracks ad 60 kids of crack free cases. Also figure 4 sows soe exaples of RGC operatio. Table RGC output for alligator cracks Crack Type Aerage of RGC Outputs ( a ) Alligator Cracks 5. Block Cracks 9.2 Logitudial 2 Traserse 2.3 Crack Free. I our experiets as sow i table, it is foud tat te output of RGC for class A (logitudial, traserse ad oters) is alost saller ta 4 ad it is bigger ta 4 for class B(alligator ad block crack). Tis criterio at te first ode of ierarcical algorit wic is sow i fig. is used. I our proposed etod te RGC is used twice. Oce at te first ode to assig te iput

6 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) iage ito oe of te class A or class B ad oce for deteriig tat te iage wic as bee assiged to class B is a alligator crack or a block crack. (a) (b) (c) Fig. 4 RGC of tree kids of cracks a) alligator, b) block ad c) logitudial I tis paper, for WSF Daubecies oter waelet D4 is adopted for iage decopositio. Aog all kids of oter waelets, Daubecies waelets are proe to be good for iage aalysis ad sytesis because of teir copact support, ore cotiuous deriaties, ad zero itegral of oter waelets [8]. Terefore, Daubecies waelets are cose i our approac ad te we are applied differet orders of Daubecies faily ad we experietally foud tat te 4 t order is te best coice for our classificatio proble wic leads to ig classificatio accuracy. Te questio arises ere is tat ow to deterie te uber of decopositio leel. To aswer tis questio it is iportat to etio tat too large te uber of ulti resolutio leels lead to loss of iforatio ad icrease te processig tie, wereas too sall te uber of ulti resolutio leels caot sufficietly effectie ad cause to decreased te accuracy of classificatio. For tese reasos i our proposed etod te axiu decopositio leel is tree ad sub bad 3 is foud to gie te best results because at tis leel udesired sall pieces, wic are probably be oise, are eliiated. Naturally, we beefit by giig a iger weigt to tis sub bad. To deostrate te discriiatio power of WSFC at sub bad 3, firstly, we fed a logitudial crack wic to te WSFC wic used oly oe decopositio leel. Ufortuately it isclassified as a traserse crack. Now to reedy tis proble, we are also used te 3 rd decopositio leel ad look at te result. Fortuately, te classificatio is successful. It is te preferece of te WSFC rater ta te istogra etod. Table sows te correspodig ea alues µ ad µ for crack free types are ery sall (less ta ). I class A ertical ad orizotal istogra ae o sigificat differeces, but ertical ad orizotal proxiity ca be used for recogitio. For class B te table sows, all paraeters for alligator cracks are bigger ta correspodig paraeters of block cracks. It is obious tere is a sigificat differece betwee ertical proxiity for alligator ad block cracks.

7 Proceedigs of te 5t WSAS It. Cof. o SIGNAL, SPC ad IMAG PROCSSING, Corfu, Greece, August 7-9, 2005 (pp ) secods te progra classifies eac iage. As te algorit is easy, ardware ipleetatio is easy wit low cost. Ackowledgeet Autors express teir taks to researc affair i Sarood Uiersity of Tecology for teir fiatial support of tis researc. 4 Coclusios Te proposed etod classified crack free iages wit 00% accuracy. Te crack iages are classified to traserse, logitudial, block ad alligators wit 98%, 90%, 89% ad 88% accuracy, respectiely. Tis etod is fast, i Refereces: []. D. Ceg et all, Noel Approac to Paeet Crackig Detectio Based o Fuzzy Set Teory, oural of coputig i ciil gieerig, October 999. [2] C.. Cao ad F. P. Ceg, Fuzzy Patter Recogitio Model for Diagosig Cracks i RC Strutures, oural of coputig i ciil gieerig, ol. 2, No. 2, April 998. [3] A C. eat et. all, Modelig Logitudial, corer ad Traserse Crackig i oit Cocrete Paeets, Iteratioal ourals of Paeet gieerig, ol. 4 (), Marc [4] T. Toikawa, A Study of Road Crack Detectio by Meta-Geetic Algorit, I AFRICON, ol., 28 Sept. - Oct, 999. [5] C. Sceffy ad. Diaz, Aspalt Cocrete Fatigue Crack Moitorig ad Aalysis Usig Digital Iage Aalysis Teciques, Iteratioal Coferece o Accelerated Paeet Testig, Reo, Neada, 8-20 October, 999. [6] B. Rai ad M.. ag, Fuzzy Logic Based Caracter Recogizer, Proceedigs of te Coputig Sciece Cogress (PCSC),2000. [7] A. K. ai, Fudaetals of Digital IageProcessig, Pretice all, 989. [8]. Ze et al,cotet-based Iage Idexig ad Searcig Usig Daubecies Waelets, Iteratioal oural of Digital Libraries, P3-328, 997.