A neurofuzzy color image segmentation method for wood surface defect detection

Transcription

1 neurofuzzy color image segmenaion mehod for wood surface defec deecion Gonzalo. Ruz Pablo. Esévez Claudio. Perez bsrac crucial sep in developing auomaed visual inspecion sysems for wood boards is image segmenaion, which aims o achieve a high defec deecion rae wih a low false posiive rae (clear wood areas idenified as defec areas). In his sudy, a neurofuzzy color image segmenaion mehod for wood surface defec deecion is proposed. The mehod is called fuzzy min-max neural nework for image segmenaion (FMMIS). The FMMIS mehod grows boxes from a se of pixels called seeds, o find he minimum bounded recangle (MBR) for each defec presen in he wood board image. n auomaic mehod o selec seeds from defecive regions as saring poins o FMMIS is also presened. The FMMIS mehod was applied o a se of 900 images of radiaa pine boards, which included samples from he following 10 caegories of defecs: birdseye and freckle, bark and pich pockes, wane, splis, blue sain, sain, pih, dead knos, live knos, and holes. The FMMIS achieved a defec deecion rae of 95 percen on he es se, wih only 6 percen of false posiives. To measure he qualiy of segmenaion, he area recogniion rae (RR) crierion was compued using as a reference he manually placed MBR for each defec. The RR achieved 94.4 percen on he es se. lso a relaive index was used o compare he qualiy of segmenaion beween FMMIS and he segmenaion module of a previously developed sysem, based on hisogram hresholding. The resuls show ha FMMIS allows us o obain significan improvemens compared wih previous work. In he las decades, several auomaed visual inspecion (VI) sysems have been developed and applied o a wide range of producs, including wood (Pham and lcock 2003). VI is an auomaed form of qualiy conrol normally achieved using a camera conneced o a compuer. The VI framework includes five processing sages: image acquisiion, image enhancemen, image segmenaion, feaure exracion, and classificaion. review of VI research applied o he inspecion of wood boards concluded ha segmenaion is ofen he mos imeconsuming par of he process, and ha i usually does no locae all defecs properly. I is necessary o develop new segmenaion algorihms ha can separae all defecs from clear wood in he image (Pham and lcock 1998). lhough mos VI sysems for wood have been developed for gray-scale images (Pham and lcock 1996,1998), some researchers have used color images (Conners e al. 1985, Brunner e al. 1992, Kline e al. 1998, Funck e al. 2003, Esévez e al. 2003a). Color image segmenaion algorihms can be classified ino one or more of he following echniques (Cheng e al. 2001): hisogram hresholding, feaure space clusering, region-based approaches, edge deecion, fuzzy approaches, neural neworks, physics-based approaches, and hybrid echniques ha combine any of he echniques jus menioned. Mos color segmenaion approaches are based on gray level (monochrome) segmenaion approaches, which can be direcly applied o each componen of a color space (Cheng e al. 2001). The selecion of a color space is applicaion dependen. Brunner e al. (1992) found ha, for images of Douglas-fir veneer, here is no advanage in ransforming he red, green, and blue (RGB) color space ino oher color spaces. The auhors concluded ha only wo color parameers, The auhors are, respecively, Research ssisan and ssociae Professors, Dep. of Elecrical Eng., Univ. of Chile, Casilla 412-3, Saniago, Chile. The auhors would like o hank Research ssisan Rodrigo Flores for his valuable help wih he experimens and sofware implemenaion. This research was suppored by Conicy-Chile under gran Fondecy This paper was received for publicaion in Sepember ricle No Fores Producs Sociey Member. Fores Producs Sociey Fores Prod. J. 55(4): PRIL 2005

2 one measuring brighness and anoher chromaiciy, are required o separae defecs from clear wood. Funck e al. (2003) compared he performances of nine segmenaion algorihms on images of Douglas-fir veneer. n algorihm ha combined clusering wih region-growing echniques achieved he bes overall performance. In VI sysems for wood here is usually a rade-off beween he defec deecion rae (rue posiives) and he rae of clear wood areas deeced as defecs (false posiives). Kline e al. (1998) found ha acual clear wood areas classified as defecs were he primary cause for yield reducion of heir prooype color VI sysem. The image scanning sysem was very sensiive o he naural variaions in he color of clear wood of red oak, and ended o idenify defecs ha were no ruly presen. Pham and lcock (1996) developed a sysem for segmening grayscale images of birch wood. The sysem consised of four modules: global adapive hresholding, muli-level hresholding, row-by-row adapive hresholding, and verical profiling. The resuls on 75 images showed a defec deecion rae of 93 percen. The sysem had difficuly disinguishing some sound knos and hard ro from clear wood areas. In a subsequen work, a pos-processing sep was performed afer segmenaion o remove false objecs and combine areas ha represen he same defec, using fuzzy logic and neural nework echniques (Pham and lcock 2003). In anoher sudy, a low-cos color VI sysem for classificaion of defecs in radiaa pine boards was developed (Esévez e al. 2003a). The image segmenaion was performed by hisogram-based muliple hresholding. The defec deecion rae achieved was 95 percen. This high rae of defec deecion was achieved a he expense of increasing he rae of false-posiives, i.e., dark grain lines segmened as defecs. One conclusion of ha sudy was he need o enhance he segmenaion process. rificial neural neworks have been widely applied o paern recogniion asks. survey on image processing wih neural neworks repored several ypes of neural neworks ha have been applied o perform image segmenaion: mulilayer percepron, self-organizing maps, Hopfield neworks, probabilisic neural neworks, radial basis funcion neworks, cellular neural neworks, consrain saisfacion neworks, and RM-based neural neworks (Egmon- Peersen e al. 2002). On he oher hand, fuzzy se heory provides a mechanism o represen and manipulae uncerainy and ambiguiy. Fuzzy operaors, properies, mahemaics, and inference rules (if-hen rules) have found considerable applicaions in image segmenaion (Cheng e al. 2001). The flexibiliy of fuzzy ses and he compuaional efficiency of neural neworks have caused a grea amoun of ineres in he combinaion of boh echniques. mong he neurofuzzy approaches, Simpson (1993) inroduced he fuzzy min-max (FMM) clusering neural nework, where clusers are represened as hyperboxes in he n-dimensional paern space. The fuzzy se hyperboxes are defined by pairs of minmax poins, and a membership funcion is defined wih respec o hese poins. The learning algorihm is a hree-sep expansion-conracion process, which has he abiliy o learn online and in a single pass hrough he daa. In his work, we propose a color image segmenaion mehod based on FMM neural neworks. The new mehod is called fuzzy min-max neural nework for image segmenaion (FMMIS). The firs sep of he mehod is he auomaic selecion of saring pixels from defecive regions. Wih his aim, a hisogrambased sudy of he color inensiies from defecive regions and grain line regions of radiaa pine boards is performed. In he second sep of he mehod, recangular boxes are grown from he iniial se of pixels wih he objecive of enclosing he defecive regions. The performance of he FMMIS mehod on he es se of pine board images is measured using he following crieria: confusion marix, area recogniion rae (RR), average processing ime, and segmenaion qualiy. The relaive ulimae measuremen accuracy (RUM) index (Zhang 1996) is used o compare he qualiy of segmenaion beween FMMIS and he segmenaion module of our previously developed VI sysem (Esévez e al. 2003a). Preliminary work on FMMIS applied o he segmenaion of knos has been repored elsewhere (Esévez e al. 2003b). Mehods Wood image daabase daa se of 900 color images (320 by 240 pixels) of radiaa pine (Pinus radiaa D. Don) boards was drawn from he Universiy of Chile daabase (Esévez e al. 2003a). The imaging sysem consised of a Naional Television Sandards Commiee (NTSC) color video camera, a frame grabber from Imaging Technology, and a 333-MHz PC Penium-II, 128 MB RM. Lighing was a mixure of fronal halogen lighs and fluorescen lamps (see Esévez e al. 2003a for deails). Spaial resoluion was 1.46 pixels/mm in boh cross-board and down-board direcions. Each image was manually labeled according o is larges defec, ino one of he following 10 defec caegories: birdseye and freckle, bark and pich pockes, wane, splis, sain, blue sain, pih, dead knos, live knos, and holes. The daa se, which corresponded o 90 images per caegory, was pariioned ino wo ses: 600 images for he raining se and 300 images for he es se. The raining se was used o adjus he FMMIS parameers. lso, 200 images from he raining se were used o make he hisogram-based color inensiy sudy wih samples aken from defecive regions and clear wood regions. The performances of FMMIS and he segmenaion module of our previous VI sysem were measured on he es se. Color of defecs and grain lines Experimens were carried ou wih 200 images of radiaa pine boards, 20 for each defec caegory. For each image, samples were aken from windows of pixels belonging o defecive regions and o grain lines. This was performed by manually placing he windows inside he regions of ineres, from where he color inensiy levels for he hree RGB color channels were recorded. The window size used depended on he size of he objec being analyzed. Once he sampling of he 20 boards for each caegory was finished, hisograms were buil for he inensiies of pixels from defecive regions and grain line areas, for each caegory and for each color channel. Because variable window sizes were used, he numbers of pixels from defecive and grain line areas were no necessarily equal in each caegory. To make he hisograms, he number of samples per caegory was chosen as he minimum value beween he number of pixels from defecs and he number of pixels from grain lines. The aim of his sudy was o obain a range of color inensiies ha allow us o selec pixels from defecive regions, FOREST PRODUCTS JOURNL Vol. 55, No. 4 53

3 as saring poins of he proposed segmenaion mehod. Neurofuzzy color image segmenaion mehod The firs sep of he proposed segmenaion mehod is o auomaically locae iniial pixels, called seeds, wihin he defecive regions. Once he seeds are deermined, hey become he inpu daa for FMMIS. The seed locaions in he image are deermined by an adapive hresholding mehod, which is based on cerain feaures of each wood board image. This allows us o ake ino accoun he grea variabiliy of he wood s color. The feaures used are he mean color inensiy value, µ, and he minimum color inensiy value, η, of he image for each channel (=R,G,B). For each color channel of he board image, a cumulaive hisogram, H, is consruced as follows: H () n = h () i n i= η for each = R,G,B [1] where n is he inensiy level (0 n 255) and h is he hisogram of he board image for channel. Since h (i) =0for all i <η, he sum in Equaion [1] sars from i =. From he cumulaive hisogram an adapive inensiy level is defined as: θ = α H( µ ) [2] where 0 α 1is a user-defined value. Typically α 0.01, since only a few pixels belonging o defecive regions are searched for as seeds, and usually he defec areas cover less han 10 percen of he image. If he color inensiies were consrained o be beween η and θ, only he darker defecs would be deeced. Bu usually here are several defecs on he same board, some brigher han ohers. To ake ino accoun his fac, an exra color inensiy level, ξ,is defined as: θ + µ ξ = [3] 2 For each board, he seeds are aken from he following inensiy range: [ η θ] ξ if θ < λ I = [4] [ η θ] if θ λ where λ is a user-defined hreshold for each channel. The raionale underlying his parameer is ha when all he defecs in a board image are no oo dark, he addiion of an exra color inensiy level (ξ ) is no useful, and i may conribue o he deecion of grain lines. The second sep is he FMMIS mehod, which places hyperboxes defined in he 2D geomeric space by pairs of minmax poins for each spaial coordinae of he image (recangular boxes in he case of 2D images). Each hyperbox fuzzy se has an associaed membership funcion ha describes he degree of membership (spaial proximiy) of a given pixel o a hyperbox in he [0,1] inerval. Seeds conained wihin a hyperbox have full membership value, and he more separaed hey are from he min-max bounds of he hyperbox, he lower heir membership values. When an inpu paern (new seed) is presened, he hyperbox wih he highes degree of membership is found and expanded o enclose he inpu paern. The hyperbox expansion is acceped only if he region conained by he expanded hyperbox is similar in color o he region enclosed by he hyperbox before he expansion. fuzzy color homogeneiy crierion is defined o compare he color similariy of wo hyperboxes. This is based on a Z-funcion (Cheng e al. 2002) of he Euclidean disance of he mean color inensiies of he wo hyperboxes, measured in he RGB space. user-defined parameer τ is inroduced o conrol he required degree of color homogeneiy for expanding hyperboxes. If he expansion crierion is no saisfied, a new hyperbox is creaed. n overlap es and a hyperbox conracion process are used o eliminae any overlaps formed during he consrucion of he hyperboxes. fer a single pass hrough all he seeds, here is a fine-uning hyperbox expansion process, which allows he hyperbox o grow if necessary unil he defec is compleely enclosed. The las sage is a hyperbox fusion process ha merges hyperboxes belonging o he same defec, o ensure ha each defec presen in he image is conained by only one hyperbox. membership funcion is used o measure he degree of spaial proximiy and color similariy beween wo hyperboxes. If he membership value is greaer han a given hreshold D, he hyperboxes are merged. For more deails abou he FMMIS mehod, refer o he research by Ruz (2003). Experimenal procedure The parameers of he adapive hresholding mehod for he locaion of seeds as well as he parameers of FMMIS were adjused using he raining se. The seeds were ordered in a vecor by raversing he image from lef o righ, op o boom. When more han 100 seeds per image were obained, his number was cu in half by aking one every oher componen in he seed vecor. To esimae he bes value of he hreshold α in Equaion [2], a receiver operaing characerisic (ROC) curve was made. On an ROC graph, he rue posiive rae is ploed in he y-axis and he false posiive rae is ploed in he x-axis, depicing he radeoff beween boh raes. Values of α generaing a rue posiive (defec deecion) rae higher han 90 percen and a false posiive (clear wood areas idenified as defec areas) rae lower han 10 percen were searched for. The parameers of FMMIS were se o (see Ruz 2003 for deails) γ = 1 (sensiiviy parameer), τ = 0.99 (degree of color homogeneiy used in hyperbox expansion), u R = 195 (fine-uning hyperbox expansion parameer), and D = 0.95 (hyperbox merging parameer). Only he red channel was used in he hyperbox fine-uning expansion process since i performed bes a separaing defecs from clear wood. This finding agrees wih ha of Brunner e al. (1992), who concluded ha knos required only a measure of brighness for image analysis (red in RGB space). To avoid noisy inpus, isolaed seeds having no neighboring seeds wihin a window of pixels were eliminaed. The performance of he FMMIS was measured on he es se using he following crieria: confusion marix, RR, RUM index of segmenaion qualiy, and average processing ime. For each of he 10 defec caegories, a confusion marix was buil as shown in Table 1. The rue posiives (TP) are defined as he number of defecs conained by hyperboxes, i.e., he number of defecs correcly deeced. The false negaives (FN) are he number of defecs ha are no conained by hyperboxes, i.e., he number of non-deeced defecs. The false posiives (FP) are he number of Table 1. Confusion marix. a Objec\hyperbox Conained No conained Defec TP FN Grain line FP TN a TP = rue posiives; FN = false negaives; FP = false posiives; TN = rue negaives. 54 PRIL 2005

4 grain lines conained by hyperboxes, i.e., he number of grain lines deeced as defecs. The rue negaives (TN) are he number of grain lines no conained by hyperboxes, i.e., he number of nondeeced grain lines. The RR was used o measure he segmenaion qualiy of he FMMIS. s a reference area, he minimum bounding recangle (MBR) was manually adjused for each defec presen in he es se. The MBR is he smalles recangle ha conains all he pixels of a defec. Because mos defecs are no recangular, he MBR conains all he pixels ha belong o he defec plus some clear wood pixels. The RR crierion akes ino accoun ha no necessarily all he pixels conained by a hyperbox belong o defecive regions. This crierion allows us o compare he area of he hyperbox buil auomaically by FMMIS wih he area of he manually placed MBR. The RR is defined as: ukp unp RR = 1 p p 100% [5] Figure 1. Hisograms from defecive regions (ligh line) and clear wood regions (dark line) for he red channel. The samples correspond o wood boards conaining he following defec caegories: a) birdseye and freckle; b) blue sain; c) sain; and d) live knos. where p is he oal number of pixels in he MBR; ukp is he number of unrecognized defec pixels, i.e., he absolue difference beween he defec pixels conained wihin he MBR and he defec pixels conained wihin he hyperbox deermined by he FMMIS mehod; unp is he number of unrecognized non-defecive pixels, i.e., he absolue difference beween he clear wood pixels conained wihin he MBR and he clear wood pixels conained wihin he hyperbox deermined by he FMMIS mehod. The performances of FMMIS and he segmenaion module of our previously developed VI sysem (Esévez e al. 2003a) were compared on he es se. Confusion marices like he one shown in Table 1 were consruced for boh mehods, bu using he crierion of wheher a defec was deeced or no. To compare he segmenaion qualiy of boh mehods, he RUM index for he percen area of he defec correcly segmened, was compued as: RUM R = S R 100% [6] where R denoes he area obained from a reference image and S denoes he area measured on he segmened image. The values of RUM are inversely proporional o he qualiy of he segmenaion resuls: he smaller he values, he beer he qualiy regarding he feaure used. The hree described measures were made considering he 10 larges objecs for each image. This consrain only affecs he birdseye and freckle caegory ha usually have more han 10 objecs per image. The seeded region growing (SRG) algorihm proposed by dams and Bischof (1994) was implemened for comparison purposes. For he wood applicaion, an adapive hresholding mehod similar o ha described by Equaions [1] and [2] was used o locae seeds auomaically. The neurofuzzy color image segmenaion mehod and he SRG algorihm were implemened in MTLB 6.5 on a PC Penium IV, 2.4 GHz, 512 MB RM. The average processing ime was measured for each image, saring from he seed selecion process and following wih he respecive segmenaion mehod. Resuls s menioned in he mehods secion, 200 images (20 per defec caegory) from he raining se were used o perform a hisogram-based color inensiy sudy. Figure 1 shows he resuling hisograms for he red channel, where he dark lines represen he hisograms generaed from clear wood pixels and he ligh lines represen he hisograms generaed from defec pixels. The hisograms overlap for mos defec caegories, excep for he spli and hole caegories. The caegories ha showed greaer overlap wih clear wood were birdseye and freckle (Fig. 1a), blue sain (Fig. 1b), sain (Fig. 1c), and live knos (Fig. 1d). In general, he blue channel presened more overlapping han he oher channels. For his reason, only he red and green channels were used o selec saring seeds for he FMMIS mehod. The parameers of he adapive hresholding mehod were se o λ R = 175 and λ G = 130 in Equaion [4]. To se he hreshold α in Equaion [2], he ROC curve shown in Figure 2 was ploed, using five differen values of α. The bes radeoff was obained for α = 0.007, reaching a TP rae of 95 percen and an FP rae of 5 percen on a subse of he raining se. On average, he number of seleced seeds per image was abou 100, i.e., 0.1 percen of he oal number of pixels of an image. Figure 3 illusraes he sep-by-sep applicaion of FMMIS o an image of a radiaa pine board. Figure 3a shows an image ha conains a dead kno as a principal objec and wo pockes as secondary defecs in he upper lef par of he image. Figure 3b shows he seeds as whie dos locaed wihin he defecive regions. Figure 3c shows hree recangular boxes deermined by he FMMIS mehod, afer a single pass hrough all he seeds. Figure 3d shows he hree boxes FOREST PRODUCTS JOURNL Vol. 55, No. 4 55

5 Figure 2. ROC curve ploing TP rae versus FP rae for five differen values of he parameer α, as shown on he curve. The bes radeoff is obained for α = Figure 3. Example showing he sep-by-sep applicaion of he proposed color image segmenaion mehod: a) original image conaining a dead kno a he cener and wo pockes a he upper lef; b) seeds are shown as whie dos locaed wihin he defecive region; c) hree boxes creaed by FMMIS afer a single pass hrough all seeds; and d) he same hree boxes afer he fine-uning expansion process. afer he fine-uning expansion process. Figure 4 shows he recangular boxes creaed by FMMIS on image samples of each of he 10 defec caegories considered. The FMMIS global and per caegory performances on he es se are summarized in Table 2. The TP rae achieved 95 percen of he oal defecs presen in he es se wih 10 defec caegories, while he FP rae corresponded o 6 percen of he oal grain lines presen in he es se. The global area recogniion rae achieved was 94.4 percen. The bes resuls were obained for he pocke, pih, dead kno, live kno, and hole caegories, which presened a TP rae higher han 95 percen, an FP rae lower han 5 percen, and an RR higher han 95 percen. The wane caegory achieved a perfec TP rae and RR rae bu an FP rae of 10.2 percen. Likewise, he spli caegory achieved a TP rae of 100 percen, an RR of 90.6 percen, and an FP level of 11.8 percen. The wors performances were obained for he birdseye and freckle, sain, and blue sain caegories, where he TP rae was lower han 95 percen, he FP rae was higher han 5 percen, and he RR was lower han 95 percen. In conras, he global performance of he segmenaion module of our previous VI sysem, which uses hisogram-based muliple hresholding, obained a TP rae of 94 percen wih an FP rae of 32 percen on he same es se. None of he caegories achieved an FP rae lower han 10 percen. The RUM crierion was compued o compare he segmenaion qualiy of he FMMIS and he segmenaion module of our previous VI sysem. Figure 5 shows he mean value of he RUM index, using 30 images per caegory. For all defec caegories, he mean values of he RUM index for FMMIS were smaller (beer) han hose for he segmenaion module of our previous VI sysem. paired -es showed ha here are significan saisical differences for all caegories (p-value lower han 0.05), excep for he birdseye and freckle and spli caegories (p-value higher han 0.05). The las column of Table 2 shows he average processing ime for FMMIS, including he seed selecion process, which reached 0.11±0.04 seconds per image. In comparison, he SRG algorihm of dams and Bischof (1994) obained an average processing ime per image of 3.03±1.30 seconds, on he same es se, i.e., FMMIS was 27 imes faser han SRG. Moreover, he SRG algorihm achieved a poor segmenaion on he birdseye, sain, blue sain, and spli caegories (TP < 50%). Only he pocke, wane, and pih caegories achieved a TP rae higher han 90 percen and an FP rae lower han 6 percen. Discussion The segmenaion module of he VI sysem previously developed obained a TP rae of 94 percen and an FP rae of 32 percen. The las figure should be compared wih he FP rae of 6 percen achieved by FMMIS. This FP rae may be furher reduced by filering ou boxes conaining only clear wood regions. The RUM crierion showed ha FMMIS segmened a larger proporion of he acual area of he defec han he segmenaion module of our previous VI sysem, for all caegories excep for he birdseye and freckle and spli caegories, which are saisically indisinguishable concerning he area of he segmened objecs. The beer segmenaion qualiy of FMMIS is due o is abiliy o segmen he complee area of he objec insead of performing parial segmenaion. Table 2 shows ha he wors FP rae performance corresponded o he spli caegory. This is due o he consrain o grow boxes along he main reference sysem, while many splis are diagonal. For diagonal splis, he boxes would cover a large non-defecive region, as can be seen in Figure 4g. possible 56 PRIL 2005

6 Table 2. Performance of FMMIS on he es se. Defec caegory TP FN FP TN TP FP RR Time (%) (sec.) Birdseye and freckle Bark and pich pockes Wane Splis Sain Blue sain Pih Dead knos Live knos Holes Global (10 caegories) ± 0.04 Figure 4. Boxes deermined by FMMIS on image samples for each of he 10 defec caegories: a) birdseye; b) blue sain; c) pocke; d) pih; e) wane; f) dead kno; g) spli; h) live kno; i) sain; and j) hole. way of dealing wih his problem is o add a box roaion sage, o allow a beer fi of he box over objecs ha do no follow he orienaion of he main reference sysem. The las column of Table 2 shows he average processing ime per caegory. The slowes segmenaion imes corresponded o he sain, blue sain, wane, and pih caegories. This can be explained because he FMMIS processing ime depends on he number of seeds as well as he number of boxes creaed. Sample boards of he sain, blue sain, wane, and pih caegories have ypically he larges defecs, and herefore end o conain more seeds. Boards of he birdseye caegory have ypically he greaes number of defecs per image. The processing ime of he seed selecion mehod was abou one-fifh of he oal processing ime. Wihin FMMIS, he slowes sage corresponded o he merging process, which depends on he number of hyperboxes formed. The average processing ime could be reduced in an order of magniude by using C programming insead of Malab. Like region-growing image segmenaion echniques, he FMMIS mehod uses a few pixels as seeds o grow regions. Neverheless, FMMIS uses only he seeds o grow recangular boxes, and herefore can easily expand a box o include a new seed when a color homogeneiy crierion is saisfied. In conras, he region-growing mehods grow by appending o each seed all neighboring pixels ha have similar properies o he seed. s a consequence, FMMIS should be faser han mos region-growing mehods, in paricular we found ha FMMIS is over 25 imes faser han he SRG algorihm in he segmenaion of wood defecs. The FMMIS mehod allows us o find he MBR of each defec presen in he wood board images. This can be viewed as a coarse segmenaion, aimed a quickly locaing all he defec areas, and separaing hem from clear wood areas. Since he FMMIS algorihm is very fas (more han 25 imes faser han an alernaive region-growing mehod), here is room for a pos-processing sep if necessary. fine-uning segmenaion sage could be added o find all he pixels wihin he MBRs ha acually belong o defec areas. However, for many applicaions, he coarse segmenaion sage may be enough, since feaures exraced FOREST PRODUCTS JOURNL Vol. 55, No. 4 57

7 Figure 5. Comparison of he qualiy of segmenaion beween FMMIS and he segmenaion module of he VI sysem previously developed (Esévez e al. 2003a), under he average RUM index for he area of he segmened objecs. Leers on he x-axis correspond o one of he 10 defec caegories: birdseye (be), pocke (po), wane (wa), spli (sp), sain (s), blue sain (bs), pih (pi), dead kno (dk), live kno (lk), and hole (ho). from he MBRs could be direcly used o classify he segmened objecs ino one of he 10 defec caegories. Once hedefecs have been locaed and idenified, he cuing process a he rough mill has o follow he main reference sysem, as he edges of he MBRs do. Conclusions The proposed color image segmenaion mehod achieved a high defec deecion rae (95%) wih a low false posiive rae (6%) on images of radiaa pine boards. key par of he mehod is he auomaic selecion of seeds belonging o defecive regions, which is based on adapive hresholding. The seed selecion procedure may be easily adjused o oher kinds of wood, wih differen illuminaion sysems, by analyzing he hisograms of wood samples. The resuls show ha significan improvemens have been obained, in comparison wih previous work, regarding he isolaion of defecs from clear wood and he qualiy of he segmenaion of defecs on images of radiaa pine boards. Lieraure cied dams R. and L. Bischof Seeded region growing. IEEE Trans. Paern nal. Machine Inell. 16(6): Brunner, C.C.,.G. Marisany, D.. Buler, D. VanLeeuwen, and J.W. Funck n evaluaion of color spaces for deecing defecs in Douglas-fir veneer. Indusrial Merology 2: Cheng, H.D., X.H. Jiang, and J. Wang Color image segmenaion based on homogram hresholding and region merging. Paern Recogniion 35: ,, Y. Sun, and J. Wang Color image segmenaion: dvances and prospecs. Paern Recogniion 34: Conners R.W., C.W. McMillin, and C.N. Ng The uiliy of color informaion in locaion and idenificaion of defecs in surfaced hardwood lumber. In: Proc. of he Firs Iner. Conf. on Scanning Technology in Sawmilling. Miller-Freeman Publicaions, Fores Prod. Soc., Madison, WI. pp. XVIII Egmon-Peersen, M., D. de Ridder, and H. Handels Image processing wih neural neworks - review. Paern Recogniion 35: Esévez, P.., C.. Perez, and E. Goles. 2003a. Geneic inpu selecion o a neural classifier for defec classificaion of radiaa pine boards. Fores Prod. J. 53(7/8):87-94., G.. Ruz, and C.. Perez. 2003b. Fuzzy min-max neural nework for image segmenaion. In: Proc. of he 7h Join Conf. on Informaion Sci. (JCIS 2003). ssoc. for Inelligence Machinery, Durham, NC. pp Funck, J.W., Y. Zhong, D.. Buler, C.C. Brunner, and J. P. Forrer Image segmenaion algorihms applied o wood defec deecion. Compu. Elecron. gr. 41(1-3): Kline, D.E.,. Widoyoko, J.K. Wiedenbeck, and P.. raman Performance of color camera machine vision in auomaed furniure rough mill sysems. Fores Prod. J. 48(3): Pham, D.T. and R.J. lcock uomaic deecion of defecs on birch wood boards. In: Proc. Insiuion of Mechanical Engineers, Par E. J. of Process Mechanical Engineering 210: and uomaed grading and defec deecion: review. Fores Prod. J. 48(4): and Smar Inspecion Sysems. cademic Press, London. 218 pp. Ruz, G New color image segmenaion mehod based on fuzzy min-max neural neworks. MS hesis. Dep. of Elecrical Engineering, Univ. of Chile, Saniago, Chile. 88 pp. (in Spanish). Simpson, P.K Fuzzy min-max neural neworks. Par 2. Clusering. IEEE Trans. Fuzzy Ses 1: Zhang, Y.J survey on evaluaion mehods for image segmenaion. Paern Recogniion 29(8): PRIL 2005