Key-Words: Feature extraction, Mathematical morphology, Template matching, Land consolidation, Remote sensing

Transcription

1 Feaure Exracion Mehod for Land Consolidaion from High Resoluion Imagery RUI GUO 1,, DAOLIANG LI 1. College of Informaion and Elecrical Engineering, China Agriculural Universiy, 17 Tsinghua Eas Road, Beijing 183, P.R. CHINA;. Key Laboraory of Modern Precision Agriculure Sysem Inegraion, Minisry of Educaion, P.O. Box 9, Beijing, 183, P.R. CHINA * Corresponding Auhor: Address: li_daoliang@yahoo.com or dliangl@cau.edu.cn, Tel: ; Fax: ,,* Absrac: Land consolidaion is a ool for increasing he area of he arable land and improving he effeciveness of land culivaion. Wih he developmen of high resoluion image, he progress of land consolidaion projec can be moniored by acquiring informaion from he image objecively. This paper presens a mehod o exrac he wells and roads in land consolidaion projec from high resoluion images. The well exracion mehod is based on he gray-level emplae maching algorihm. The road exracion mehod is based on mahemaical morphology, which is a mehod for deecing image componens ha are useful for represenaion and descripion. The vecor planning maps and high resoluion images used o monior he compleion of land consolidaion projec are regisered. The candidae areas are creaed using he funcions of buffer and exracion by mask in GIS. The well emplae is seleced manually from he image. The emplae is used o find he wells which mach he emplae perfecly. In he road exracion secion, Top-ha ransform and gray dilaion are used o filer he noise of he image. In his way he road feaure in he image became wider and even more obvious o be recognized. Then image binarizaion and hinning algorihm are used o exrac he one-pixel cenerline of he road. A las, he hinning resuls are convered o he final vecor deecion resuls. Key-Words: Feaure exracion, Mahemaical morphology, Templae maching, Land consolidaion, Remoe sensing 1 Inroducion Land consolidaion is a ool for improving he effeciveness of land culivaion, land produciviy and also he oal facor produciviy if i induces and enhances echnical progress and increases scale economies. Consolidaion deals wih a large number of phenomena, such as fields, roads, and land use, all of which exhibi characerisic forms and paerns which can be analyzed as o heir exising spaial organizaion, or as o heir changing spaial organizaion hrough ime [1]. Wih he developmen of high resoluion imagery and he improved feaure deecion mehod, remoe sensing has become an effecive means of land use monioring. There are wo kinds of algorihm of image maching including gray emplae maching and feaure maching. The gray emplae maching is abou 1 dimension or dimension sliding emplae which is mainly based on space domain. The differences in his algorihm are he seleced emplae and rules. The feaure maching is mainly abou exracing poins, lines and regions as he bases from he image and hen maching. Toally, he gray emplae maching needs more calculaion bu more accurae []. Loi and Giraudon[3,4] used a correlaion based algorihm wih an adapive window-size ha is consrained by an edge map exraced from he image. They presened resuls on real aerial images. Inille and Bobick[5] presened a sereo algorihm ha incorporaes he deecion of he occlusion regions direcly ino he maching process. They developed a dynamic programming soluion ha obeys he occlusion and ordering consrains o find a bes pah hrough he dispariy-space image. They also used ground conrol poins o eliminae sensiiviy o occlusion cos. Xiong e al [6] presened a sereo maching approach which inegraes area-based and feaure-based processes. Fua[7] described a correlaion based muliresoluion algorihm which is followed by inerpolaion. Anandan[8] described a hierarchical compuaional frame work for he deerminaion of dense moion fields from a pair of images. A number of researchers have used dynamic programming o solve globallv he maching problem [9, 1, 11, 1, and 13]. The exising road deecion approaches cover a wide variey of sraegies, using differen resoluion aerial or saellie images. The approaches are generally ISSN: Issue 3, Volume 6, March 9

2 classified according o he degree of auomaion: auomaic mehod and semi-auomaic mehod. Semi-auomaic sysems assised by an operaor seem o be more effecive for road deecion now. However, he developmen rend of road deecion is auomaic deecion sysem in he near fuure. Afer deecing he road nework, here may be a few wrong resuls, which can also be correced by he sysem operaor. Auomaic mehods usually deec reliable hypoheses for road segmens hrough edge and line deecion and hen esablish connecions beween road segmens o form road nework [14]. Hinz e al. [15] inegrae deailed knowledge abou roads and heir conex using explicily formulaed scale-dependen models. The knowledge abou how and when cerain pars of he road and conex model are opimally exploied is expressed by an deecion sraegy. Mokharzade e al. [16] reas he possibiliy of using arificial neural neworks for road deecion from high-resoluion saellie images on a par of RGB IKONOS and Quick-Bird images from Kish Island and Bushehr Harbor, respecively. Aemps are also made o verify he impacs of differen inpu parameers on nework s abiliy o find ou opimum inpu vecor for he problem. A variey of nework srucures wih differen ieraion imes are used o deermine he bes nework srucure and erminaion condiion in raining sage. Lapev e al. [17] proposed a new approach for auomaic road deecion from aerial imagery wih a model and a sraegy mainly based on he muliscale deecion of roads in combinaion wih geomery-consrained edge deecion using snakes. I allows for he firs ime a bridging of shadows and parially occluded areas using he heavily disurbed evidence in he image. Semi-auomaic mehods require operaor o provide some informaion o conrol he deecion ineracively. Mos semiauomaic approaches search for an opimal pah beween a few given poins. Gruen and Li e al. [18] and Merle e al. [19] connec poins using dynamic programming, model-driven linear feaure deecion algorihm based on dynamic programming. Gruen e al. [] developed linear feaure deecion mehod using dynamic programming and LSB-Snakes. They combined characerisics of snakes and Adapive Leas Squares Correlaion mehod. This mehod migh need large compuaion ime on highresoluion images because of is linear sysems. Park and Kim [1] presened a road deecion algorihm using emplae maching. Bu he limiaion is ha i requires iniial seed poins on he road cenral lines and each road segmen requires separae seed poins. Shukla s scheme is based on he cos minimizaion echnique []. The cos is esimaed by aking various facors ino consideraion such as variance, direcion, lengh and widh of he road. The process sars wih he selecion of seed poins provided by he user. The approach is called as pah following as i follows he pah having minimum cos repeiively. Thus he pah having minimum cos will be considered as a par of he road. Dal Poz e al. [3] presened a dynamic programming approach for semi-auomaed road deecion from medium- and high-resoluion images, proposing a modificaion of meri funcion of he original dynamic programming approach, which is carried ou by a consrain funcion embedding road edge properies. The aim of his paper is o develop an inegrae scheme o exrac he feaures o be moniored in land consolidaion projec which includes a mehod of well exracion based on emplae maching and a mehod of road deecion mehod based on mahemaical morphology in land consolidaion area. The paper begins wih a brief overview of characerisics of he sudy area. Afer ha he well exracion mehod is illusraed in deail. This is followed by he road deecion flow and he deail algorihm. The algorihm is implemened in he Malab ools. Resuls achieved wih high resoluion saellie image are depiced. Finally, conclusions are drawn. Sudy Area The sudy area is locaed a N longiude and 4 8 E laiude in he norheas of Beijing (Fig. 1) and is influenced in some pars by differen ypes of land improvemen works, such as reclamaion, waercourse recificaion and land consolidaion works, which have a large impac on he rural landscape. This area is characerized by rolling opography and flourishing vegeaion. One Quick Bird image, acquired on June 6, wihou any clouds/hazes, is used in his sudy. A Quick Bird image has four muli-specral bands (i.e. nearinfrared (NIR), red, green, and blue) wih.44- mere spaial resoluion and one panchromaic band wih.61-mere spaial resoluion. Boh daes of imagery were geo-referenced o a Transverse Mercaor projecion and Krasovsky spheroid wih an RMSE of 1 pixel. I was necessary o radiomerically normalize he muliple daes of remoe sensor daa even hough hey were obained on near anniversary daes. ISSN: Issue 3, Volume 6, March 9

3 ShunYi Fig.1. Locaion of Sudy Area 3 Well Exracion Scheme 3.1 Gray-level emplae maching There are a few kinds of mehod o describe he gray-level emplae maching including Squared difference, Normalized squared difference, Cross correlaion, Normalized Cross correlaion, Correlaion coefficien, Normalized Correlaion coefficien. If f ( xy, ) represens M N as he original image, hen ( j, k ) represens J K ( J M, K N ) as he emplae. While he emplae ( j, k ) moves over he original f ( xy, ), he region below he emplae is called sub-image f ( xy, ). When he gray value of emplae ( j, k ) equals he gray value of he sub-image f ( xy, ), he sub-image is he objec which we are looking for. In pracice when he difference beween he wo images comes o a hreshold value, he sub-image is also he image. The similariy degree beween images could be described as squared difference as following: J 1 K J K SD( x, y) = f ( x, y) ( j, k) = [ f ( x + j, y + k) ( j, k) ] (1) j= k= j= k= When he SD( x, y ) is less, which means he searching image more like he emplae. The formula (1) can be normalized as Normalized squared difference: [ ( j, k) f( x+ j, y+ k) ] SDN( x, y) = () j= k= [ f( x+ j, y+ k) ] [ ( j, k) ] j= k= j= k= SDN( x, y) and SD( x, y ) can boh be used o search he objec. The formula (1) can also be deployed as [ ] SD( x, y) = f ( x + j, y + k) j= k= J 1 k 1 j= i= j= k= [ ] ( j, k) f( x+ j, y+ k) + ( j, k) (3) The hird par of formula (3) means he oal energy of he emplae, which is a consan when he emplae is ensured and is no relaive wih he ( x, y ). The firs par represens he energy of sub-image, which is changed by ( x, y ). The second par reflecs he relaiviy beween he emplae and he sub-image, which is also changed by ( x, y ).When he emplae and he sub-image are maching, he value comes o he maximum. So cross correlaion can be used o measure he relaiviy: Cxy (, ) = ( jk, ) f( x+ jy, + k) (4) j= k= This can also be normalized as: ( j, k) f( x+ j, y+ k) CN( x, y) = (5) j= k= [ f( x+ j, y+ k) ] [ ( j, k) ] j= k= j= k= The emplae described above only use he pixel informaion which impacs he accuracy. To be more accurae, formula (3) can be modified as ( )( ) Rxy (, ) = ( jk, ) f( x+ jy, + k) f (6) Where = j= k= j= k= j= k= ( j, k) f = f( x+ j, y+ k) ISSN: Issue 3, Volume 6, March 9

4 R( xy, ) is called Correlaion coefficien N x y ( ( j, k) )( f( x+ j, y+ k) f ) j= k= (, ) = (7) ( f( x+ j, y+ k) f ) ( ( j, k) ) j= k= j= k= Bu he calculaion is huge and is no calculaed easily. The oal calculaion is he following: Calculaion= RN( x, y)* SearchCoun So formula (7) could be changed o ( j, k) f( x+ j, y+ k) J* K* * f j= k= RN( x, y) = (8) f ( x+ j, y+ k) J* K* f ( j, k) J* K* and j= k= j= k= ( j, k) J* K* are cons. j= k= 3. Well exracion algorihm On he basis of he demand of land consolidaion projec, he algorihm presens a soluion ha impors he vecor map of he projec and regisered o he Quick Bird image, which means coarse maching in a radiional way. According o he vecor map, he posiions of wells o be exraced in he image would be probably esimaed. Then emplae is used o mach he wells and correlaion coefficien is calculaed o search in he buffer disance. This mehod no only combine he vecor map wih he projec, bu has been grealy reduced he calculaion of he maching operaion. The improving algorihm is as following (Fig..): 1. Load he vecor map and image of he land consolidaion projec.. Regiser he vecor map and he image o be a coarse maching. Afer he regisraion boh layers are in he same coordinae sysem. According o he coarse maching, he vecor layer of wells is creaed. A he same ime, aribues of well layer also has been creaed. 3. Selec a well as he maching emplae image in he image using he AOI ool. Then record he lengh and widh of he emplae wih all bands. If he image has only one band, hen calculae he only band in he operaion. When he emplae has more han one band hen changes he emplae image and he sub-image o gray images. Boh formulas described are perfec: ( R+ G+ B) or.99* R+.587* G+.114* B 3 Fig.. Well Exracion Scheme ISSN: Issue 3, Volume 6, March 9

5 4. Ge he posiions of he wells of vecor map and sore hem in a file. For each posiion ge a recangle buffer such as m. Then glide he emplae over he buffer region and calculae he correlaion coefficien. 5. Se a hreshold value of correlaion coefficien for each search region. If he calculae value is greaer han he hreshold hen he posiion is he well o be exraced. The hreshold value can be esimaed by wo knowable emplaes. 6. For each well ha has been mached, creae a poin layer and check he amoun and he posiions. 3.3 Resuls The reorganizaion accuracy of he algorihm is limied by he noise of he image, he regisraion accuracy of maps, he size of emplae and he correlaion coefficien. The buffer disance is *m. The resoluion of QuickBird image is.61m afer image fusion. So he real searching area is 1 * 1 m. We choose wo emplaes and wo hreshold value o make an experimen as he above algorihm. The wo emplaes are 15 * 16 pixels (emplae 1),1 * 11 pixels (emplae ) (See in Fig.3.) There are nine posiions of wells ha could exis. We use he nine posiions o mach he wells. The resuls are described as following able: The resuls lised in he able explain ha he exracion accuracy is relaive o he size of emplae and he hreshold value. When he value is lower (.7), here are no omied wells. Bu he wrong wells are found. On he conrary, when he value is higher (.8), here may be some omied wells. Bu he righ wells are more. Anoher facor is he size of he emplae. When he emplae is he same size of acual well such as1 * 11 pixels (emplae ), he accuracy is higher. When he emplae is large han he real well, he accuracy is lower. So all illusraions above are he algorihm and flow of well exracion.(see in Fig.4.) Fig.3. Two Kinds of Templaes Table 1: Well Exracion Resuls Templae Templae 1 Templae Threshold Righ 1 6 Value 1 Wrong 8 3 (.7) Omission Value (.8) Accuracy 11.1% 66.7% Rae Righ 3 7 Wrong 1 Omission 6 1 Accuracy Rae 33.3% 77.7% Fig.4. Well Exracion Resuls 4 Road Deecion Scheme 4.1Principle of Mahemaical Morphology Mahemaical Morphology is a mehod for deecing image componens ha are useful for represenaion and descripion. The echnique was originally developed by Maheron and Serra [4] a he Ecole des Mines in Paris. I is a se-heoreic mehod of image analysis providing a quaniaive descripion of geomerical srucures. A he Ecole des Mines hey were ineresed in analyzing geological daa and he srucure of maerials. Morphology can provide boundaries of objecs, heir skeleons, and heir convex hulls. I is also useful for many pre- ISSN: Issue 3, Volume 6, March 9

6 and pos-processing echniques, especially in edge hinning and pruning. Generally mos morphological operaions are based on simple expanding and shrinking operaions. The primary applicaion of morphology occurs in binary images, hough i is also used on grey level images. I can also be useful on range images [5] Se operaions These ransformaions involve he ineracion beween an image A (he objec of ineres) and a srucuring se B, called he srucuring elemen. Typically he srucuring elemen B is a circular disc in he plane, bu i can be any shape. The image and srucuring elemen ses need no be resriced o ses in he D plane, bu could be defined in 1,, 3 (or higher) dimensions Dilaion Dilaion of he objec A by he srucuring elemen B is given by A B = x Bˆ I A { : } x The resul is a new se made up of all poins generaed by obaining he reflecion of B abou is origin and hen shifing his relecion by x Erosion Erosion of he objec A by a srucuring elemen B is given by AΘ B= { x: Bx A} Two very imporan ransformaions are opening and closing. Now inuiively, dilaion expands an image objec and erosion shrinks i. Opening generally smoohs a conour in an image, breaking narrow ishmuses and eliminaing hin prorusions. Closing ends o narrow smooh secions of conours, fusing narrow breaks and long hin gulfs, eliminaing small holes, and filling gaps in conours Opening The opening of A by B, denoed by Ao B, is given by he erosion by B, followed by he dilaion by B, ha is Ao B= ( AΘB) B Closing Closing is he dual operaion of opening and is denoed by Ao B. I is produced by he dilaion of A by B, followed by he erosion by B A B= ( A B) ΘB 4. Road deecion algorihm The flow char of road deecion algorihm is described as he Fig.5. The Quick Bird image and he vecor planning map are regisered in he same geographic coordinae. The funcions of buffer and exracion by mask in GIS are used o obain he candidae area, which could be used o monior wheher he roads have been buil up or he lengh of he roads. Afer he candidae area image was creaed, he filer echnique is inroduced o emerge he roads. The filer echnique conained Top-ha ransform and dilaion algorihm. Then he hisogram of he image was creaed and hreshold could be calculaed auomaically, which is called binarize image. There migh be some parcels of oher objecs like grass, rees, houses ec. hey could be wiped off in erms of shape and area. The hinning algorihm is used and hen he roads could be convered o he vecor. According o he direcion and he lengh of a road, he iny lines are also eliminaed. Afer he seps above, he roads are deeced. Quick Bird Image Buffer Exracion Top-ha Transformaion Dilaion Image Binarizaion Delee Isolaed Area Thinning Connecion Road Cenerline Fig.5. Road deecion flow Vecor Planning Map 4..1 Generaing candidae area The planning map was edied before he projec, according which he roads should be consruced. There may be some roads ha were no consruced as plan exacly. The road buffer is creaed wih he widh of 3m, in which mos roads exis. The buffer hen is used as a mask o deec he candidae area. The candidae area is shown in Fig.6. ISSN: Issue 3, Volume 6, March 9

7 Fig.7. Opening by fla srucuring elemen [Doughery, 199] Fig.6. Candidae area 4.. Image filer In order o smooh he image noise, Top-ha ransformaion and dilaion are used o make he road more recognizable. The Top-Ha ransformaion is one of he gray scale morphologic algorihms and is beneficial in finding pixel clusers ha are ligh on a surrounding relaively dark background. I can be used o find neuronal cells in a issue sample and o deec blood vessels from an image using an X-ray sysem and fluorescen dyes [6]. This operaion is illusraed in Fig.7. The ransformaion processes original signal f wih opening by fla srucuring elemen g. Fig.8 indicaes ha he peaks are deeced as a Top-Ha by subracing an opened image form he original image. The opening operaion includes wo procedures, erosion and dilaion. Because he srucuring elemen is fla, he erosion is simplified o find he minimum gray level and he dilaion o find he maximum during process. This operaion seems o be able o deec he objecs above he ground, bu unforunaely he deecable size of objec depends on he size of he elemen. Therefore we use his opening operaion jus o eliminae he peaks, like noise in elevaion space [7]. The dilaion operaion can improve deecion qualiy and accuracy afer he Top-ha ransformaion. Fig.8. Top-Ha Transformaion [Doughery, 199] 4..3 Image Binarizaion Binarizaion of gray level images afer he preprocessing is a special case of segmenaion wih wo labels, which is he mos imporan sep of he algorihm. The purpose is simplifying he image o deec he conour. The hisogram of his kind of image is described in Fig.9. According o he hisogram characerisic, he hreshold is calculaed exacly as following [8]. From he peak of he image oward he righ side: T1() = ( p 5 + p 4 + p 3+ p + p 1+ p)/6 T() = ( p + p+ 1+ p+ + p+ 3+ p+ 4 + p+ 5)/6 p represens he numbers of poin whose gray value is. θ () = arcan[( T1() T())/6] π If T1() T() and θ (), is he 18 hreshold. There migh be no one, so he firs is jus he value. ISSN: Issue 3, Volume 6, March 9

8 Fig.9. Hisogram of image 4..4 Delee Isolaed Area There migh be a few noise parcels lef afer he binarizaion operaion, which need o be eliminaed. The mehod is described as following [9]: (1) Calculae he area a, perimeer p and shape index of each parcel. () Delee he parcel me he condiions of a1 < a< a and >.1, where r = a/ p, a means he number of pixel in a parcel and p means he number of pixel in a parcel boundary. a 1 means ha he mos area of a parcel belong o a road. a means ha he mos area of isolaed parcel. (3) Direcion dilaion Afer sep, here are some small parcel lef only, bu he road is sill no conneced. So he direcion dilaion operaion is used o connec he disconinuous segmen. (4)Delee he small parcels 4..5 Thinning Thinning is anoher mahemaical morphology operaion which can be used o deec he cenerline. The general definiion of a hinning of a se A by a srucuring elemen S is ha we remove from A a par of A specified by he hi-or-miss ransform. The hinning is denoed by A B and may be wrien in se noaion: The deails are described in Fig.1. Fig.1.Thinning algorihm 4.3 Resuls A piece of candidae area is cu as he es image in Fig.11. The seleced candidae area includes ypical land consolidaion roads. There are also some rees along wo sides of he road and he specrum of some blocks is similar wih he road. Fig.11. A road of sudy area Fig.1 shows he original RGB image has been convered o he gray image auomaically in Malab by he funcion of rgbogray. The roads presen whie lines, bu here sill are some rees shade he edges of he roads. Fig.1. RGB image o gray image Afer convering he RGB image o gray image, he roads need o be enhanced and he background needs o be weaken. Fig.13 shows he resuls afer image filer which conain he Top-ha ransformaion and dilaion. The funcions in Malab are imopha and imdilae. Fig.13. Top-ha ransformaion ISSN: Issue 3, Volume 6, March 9

9 Binarizaion is he mos imporan sep in he whole scheme because i compleely differeniaes he roads and background. I is a kind of image segmenaion, which ses a hreshold and divides he image ino wo pars. The hreshold is acquired from he formula described above. Fig.14 and Fig.15 show ha he profile of he road o be deeced has been come ou. The funcion imbw is used. Bu here are sill some gaps and redundancies. Fig.14. Image Binarizaion Fig.15. Deleing Isolaed Area Afer he connecion and hinning operaion, he final cenreline of road has been deeced and is shown in Fig.16 Fig.16. Thinning Taken ogeher, he final resuls of road deecion scheme have been carried ou wih a piece of image in ShunYi land consolidaion area and he resuls seem o be good and pracical. 6 Conclusion A mehod of feaure exracion based on gray-level emplae maching and mahemaical morphology for land consolidaion has been briefly described. The mahemaical morphology is easy o undersand and implemen. The experimen resul on Quick Bird saellie image ells ha our mehod can be one of he soluions for developing fully operaional feaure exracion sysem for land consolidaion. The high-resoluion images are preferable due o use of widh and variance informaion for road deecion. On he oher hand, he mulispecral informaion should also be fully used. The concep of semi-auomaion has shown o be excellen for pracical applicaions, as here is always an ediing opion when some auomaion fails due o low image qualiy, disurbances and oher effecs, bu he auomaic mehod should be he sudied in he feaure. Acknowledgemen The research was financially suppored by Sae science and echnology suppor projecs (Conrac No: 8BAB38B4). References: [1] Wu, Z., Liu, M., Davis, J., Land consolidaion and produciviy in Chinese household crop producion, China Economic Review, Vol 16, Issue 1, 5, pp [] Changming. Sun, Muli-Resoluion Recangular Subregioning Sereo Maching Using Fas Correlaion and Dynamic Programming Techniques, CMIS Repor No. 98/46, CSIRO Mahemaical and Informaion Sciences, Macquarie Universiy Campus, [3] J.-L.Loi, G.Giraudon, Adapive vvindovv algorihm for aerial image in Proceedings (Jerusalem, Israel), Inernaional Conference on Paern Recoaniion, vol. A, pp.71-73, IEEE Compuer Sociey Press, Ocober [4] J.-L. Loi, G.Giraudon, Correlaion algorihm wih adapive window for aerial image in sereo vision, European Symposium on Saellie Remoe Sensing(EUROPTO), Rome, Ialy, 1994, pp [5] S.S.Inile, A.F.Bobick, Dispariy-space images and large occlusion sereo, Proceedings of European Conference on Compuer Vision, Sockholm, Sweden, [6] Y.Xiong, D.Wang, G.Zhang, Inegraed mehod of sereo maching for compuer vision, SPIE Proc, Applicaion of Digial Image Processing XIX, Augues 1996, pp [7] P.Fua, Aparallel sereo algorihm ha produce dense deph maps and preserves image feaures, Machine Vision Applicaion, Vol.6, 1993, pp [8] P.Anandan, A compuaional framework and an algorihm for he measuremen of visual moion, Compuer and Informaion Science, Universiy of Massachuses a Amhers, Augus [9] S.A.Lloyd, A dynamic programming algorihm for binocular sereo vision, GEC Journal of Research, Vol.3, 1985, pp [1] G.L.Gimel farb, V.M.Kro, M.V.Grigorenko, Experimens wih summarized inensiy-based dynamic programming algorihms for reconsrucing digial errain modal, Inernaional Journal of Imaging Sysems and Technology, Vol.4, 199, pp7-1. [11] R.Baldwin, H.Yamada, K.Yamamoo, Dispariy space and dynamic programming for auomaic producion of very dense range maps, ISSN: Issue 3, Volume 6, March 9

10 Colserange PhoeGrammery Mees Machine Vision, Vol.1395 Sepember 199,pp [1] Y.Oha, T.Kanade, Sereo by inra- and inerscanline search using dynamic programming, IEEE Transacions on Paern Analysis and Machine Inelligence, Vol PAMI-7, 1985, pp ] [13] A.Rojas, A.Calvo, J.Munoz, A dense dispariy map of sereo images, Paern Recogniion Leers, Vol.18, 1997, pp [14]. Wiedemann, C., Heipke, C., Mayer, H., Hinz, S., Auomaic exracion and evaluaion of road nework from MOMS-P imagery. In. Archiv. Phoogrammery and Remoe Sensing. 3 (Par 3/1), 1998, pp [15]. Hinz, S., Baumgarner, A., Auomaic exracion of urban road neworks from muli-view aerial imagery. Journal of Phoogrammery & Remoe Sensing. Vol.58, 3, pp [16]. Mokharzade, M., Valadan Zoej, M.J., Road deecion from high-resoluion saellie images using arificial neural neworks. Inernaional Journal of Applied Earh Observaion and Geoinformaion, Vol.9, 7, pp.3-4. [17]. Lapev, I., Mayer, H., Lindeberg, T., Ecksein, W., Seger, C., Baumgarner, A., Auomaic exracion of roads from aerial images based on scale space and snakes. Machine Vision and Applicaions, Vol.1,, pp [18]. Gruen, A., Li, H., Road exracion from aerial and saellie images by dynamic programming. Journal of Phoogrammery and Remoe Sensing, Vol.5, No.8, 1995, pp.11-. [19]. Merle, N., Zerubia, J., New Prospecs in Line Deecion by Dynamic Programming. IEEE Trans Paern Anal Mach Inell, Vol.18, No.4, 1996, pp [] Gruen, A., Li, H., Semi-Auomaic Linear Feaure Exracion by Dynamic Programming and LSB-Snakes. Phoogrammeric Engineering & Remoe Sensing, Vol.63, No.8, 1997, pp [1] Park, S., Kim, T., Semi-Auomaic road Exracion algorihm from IKONOS images using emplae maching. Proc. nd Asian Conference on remoe Sensing, 1, pp []. Shukla, V., Chandrakanh, R., Ramachandran, R., Semi-auomaic road exracion algorihm for high resoluion images using pah following approach. In: Indian Conference on Compuer Vision, Graphics and Image Processing. Ahmadabad,, pp [3] Dal Poz, A. P., do Vale, G. M., Dynamic programming for semi-auomaed road exracion form medium and high resoluion images [J], Inernaional Archives of he Phoogrammery, Remoe Sensing and Spaial Informaion Sciences, 3, 34: Par 3 /W8, pp [4] J. Serra. Image Analysis and Mahemaical Morphology. Academic Press, 198. [5]hp://homepages.inf.ed.ac.uk/rbf/CVonline/LO CAL_COPIES/OWENS/LECT3/node3.hml [6] Doughery, Edward R An Inroducion o Morphological Image Processing, SPIE Opical Engineering Press, Cener for Imaging Science Rocheser Insiue of Technology. [7] Shao, Y., Chen, L. Objec Segmenaion in Elevaion Space Using Mahemaic Morphology. [8] Wang, Yaoge. Road exracion from remoe sensing image based on mahemaical morphology, MS Thesis. Dep. of Carography and Geographic Informaion Engineering, Informaion Engineering Universiy of he People's Liberaion Army. [9] An, R., Feng, X. Road Feaure Exrac ion form Remoe Sensing Classified Imagery Based on Mahemaica lm orphology and Analysis of Road Neworks. Journal of Image and Graphics, Vol.8,, pp ISSN: Issue 3, Volume 6, March 9