2.1 Ship feature selection. 2. Ship feature selection and extraction

Transcription

1 3rd Internatonal Conference on Multmeda Technology(ICMT 03) Mult-evdence fuson recognton of shp targets n sea battlefeld s remote sensng mages Yu An,, Xaofe Wang, Xuezh Xa and Ln L Abstract Relably recognzng shp targets n the sea battlefeld has become an ncreasngly pressng need, as the capabltes for mage acquston are growng rapdly In the research, a modfed shp targets fuson recognton model based on the Dempster-Shafer evdence theory s proposed The algorthm frstly detects shp targets from the sea battlefeld s remote sensng mage Then extracts the multple mage features of these target canddate areas as the evdence to recognze the shp targets Fnally, recognzes the shp targets from the mage usng the Dempster-Shafer evdence theory based on multple shp features, and sends the recognton result Experment show that ths method can be used to relably and effectvely recognze targets nformaton n the sea battlefeld Keywords: target recognton; Dempster-Shafer evdence; shp features; remote sensng mage Instructon As one of the man areas of operatons of modern war, Sea battle feld stuaton vary from mnute to mnute n battle Shp targets are the key targets of the martme montorng and wartme attack Rapd and accurate dentfcaton of shp sea battlefeld tactcs and provde support for decson-makng for commanders, whch s greatly related to the success or falure of the battle Many judgng methods of warshp target ntenton n a sea battle, they are bult on the bass of the target recognton In modern sea battles, n order to make the stuaton assessment and battlefeld decson, we should frst solve how to recognze accurately dfferent knds of battle shps n the dstance Yu An ( ) College of Computer Scence and Technology, Harbn Engneerng nversty Harbn, Chna e-mal: anyu73@6com Yu An Wuhan Dgtal Engneerng Insttute Wuhan, Chna 03 The authors - Publshed by Atlants Press 79

2 Prevous research for shp target recognton s manly based on the synthetc aperture radar or forward-lookng nfrared radar mage data, these methods are usng the dfference of Shp Superstructure to recognze the type of shp targets[,,3] In recent years, wth the development of satellte remote sensng technology, emergng a number of magng satellte wth hgh spatal resoluton and short revst perod, whch provdes abundant data source martme reconnassance and target montorng sng optcal remote sensng mage to recognze shp targets becomes possble For example, Zhang[4] use the support vector machne technology to recognze the shp targets n Quck Brd satellte data, Lan et al[5] propose a modfed Zernke moments model for recognzng the large warshp n aeral mages nlke the methods proposed n the above, focusng on the problem of low recognton rate whch s because of the shp mage target recognton wth mprecse, uncertan, ncomplete etc, a new research has been done based on the multfeature fuson recognton dea Wth resoluton better than 4m optcal remote sensng mage data as an object of study, ths paper proposes a mult-evdence fuson recognton method based on the shp features The method consst of the followng steps: Frstly, detectng and bnarzng the targets area n the mage, for a sngle feature nformaton s nsuffcent or easly falure to recognze the shp targets, n order to get enough useful complementary nformaton, extractng the length, wdth, area and many features from the targets slce as the evdence sources; Then, establshng the Dempster-Shafer(D-S) fuson recognton framework and computng the value of Basc Belef Assgnment of each feature; Fnally, usng D-S combnaton rule to complete the decson level fuson and output shp target recognton results Shp feature selecton and extracton Shp feature selecton Spatal dstrbuton characterstcs of the shp targets manly descrbe ther physcal propertes, ncludng sze, brghtness, and shape feature parameters Sze and shape features emphasze dstngush objects based on geometrc parameters, whch can reflect dfferences n the nature of shp n the mltary feld, such as frgate and arcraft carrer In order to be able to better and more accurate recognton the shp targets, frstly need to analyze the dfferent types of shp targets, then select the features whch can be easy to dstngush the shp targets In the hgh resoluton remote sensng mages, the sze and shape features of targets can be easy to 70

3 extracton In ths paper, select the followng features as the evdence for recognton of shp targets Sze features: shp length f, shp wdth f, shp aspect rato f, shp surface 3 area f 4 Shape features: shp bow f, shp stern 5 f, shp juncton 6 f 7 Shp feature extracton and defnton In order to extract the shp features from mages, we need to bnarze target mage f ( x, y) and calculate the moment features of mage The ( p + q) order orgn moment and the ( p + q) order central moments are defned as: p q m pq j f, j) = j ( () p q µ = ( x) ( j y) f (, j) () pq Where ( x, y) s the centrod coordnate of target mage f ( x, y) Normalzed central moments: j N ( x x) = = (3) 00 N ( y y) = = (4) 00 N ( x x)( y y) = = (5) 00 Target spndle drecton s defned as: arctan( *, + ( ) + 4* ) θ = (6) arctan( + ( ) + 4*,* ) Where θ s the angle between target spndle and axs, π π θ (, ) When ncludng geographc nformaton n mage, θ can be express shp headng Rotate target spndle to the horzontal and get the target mnmum boundng rectangle 7

4 f s defned as the length of boundng rectangle f s defned as the wdth of boundng rectangle f s defned as the aspect rato of boundng rectangle 3 Four knds of shp target aspect rato statstcs are as follow (Table ): Table Aspect rato tables of warshps Warshp Aspect Rato Destroy(Cruser) 85: ~ 05: Frgate 8: ~ 95: Arcraft carrer 55: ~ 8: Obvously, the dfferent uses of the shp has dfferent aspect rato nterval Therefore, t can recognze dfferent types of shp by usng ths feature, especally recognzng the arcraft carrer from the fleet f s defned as the sum of detecton target regon pxels Beng much larger 4 than other shps, arcraft carrers can be effectvely dstngushed from other shps based on f The actual area of the shp target s the sample nterval of pxel multpled by the total number of target pxels 4 Accordng to the characterstcs of modern sea battle, mltary acton manly focus on recognzng destroyers, crusers, frgates and arcraft carrers n the sea battlefeld Shp bow, body and stern have obvous geometry features For target recognton, they can be smplfed as trangle, rectangle and trapezod (fgure ) Fg Smplfed features of shp target In order to descrbe effectvely the shp target s bow and stern shape characterstcs, the target s dvded nto areas Each area sze s as close as possble to the sze of bows or sterns In the actual shp shape feature extracton, t s generally thought that mage resoluton should be better than 4m Consderng that a bow or a stern n general s about ~ 8 3 of the whole shp, the area number dvded s 3 ~ 8 To descrbe accurately the shape characterstcs of bows and sterns and ensure the algorthm effcency, the area s number s 5 and marked A A 5 as follows Fg Example of a shp slce A A A 3 A 4 A 5 As fgure shows, bows and sterns are represented by A and A In ths paper, the shapes of bows and sterns are abstracted to combnaton of trangle 5 and 7

5 trapezod In shp targets recognton, the juncton feature of shp targets s the number of junctons of bows and sterns f, 5 f and 6 f are defned as follows(table ) 7 Table Shp shape features and Juncton features Warshp Features bow( f ) stern ( 5 f ) 6 juncton( Destroy(Cruser) trangle trapezod 7 Frgate trangle trapezod 7 Arcraft carrer trapezod trapezod 8 f ) 7 3 Mult-feature fuson recognton of shp targets based n D-S evdence theory Shp target area, length, wdth, aspect rato, shape and juncton feature are ndependent of each other It s dffcult to meet target recognton requrement dependng on a sngle feature For example, a target whose shape s meet requrement s not the type ntended to dentfy but there s a possblty of belongng to the type In order to use effectvely varous features to mprove target recognton rate, D-S evdence theory s appled to fuse the nformaton from dfferent features[6] The decson module can fnd the target s type attrbute and confdence level Fgure 3 shows the framework of mult-feature fuson and recognton arthmetc Evdence D-S fuson processor Target Image Target feature extracton Evdence Evdence 3 Evdence n Basc Belef Assgnment constructon D-S nference fuson recognton Target recognton result Fg 3 The framework of mult-feature fuson and recognton arthmetc Let Θ be a unverse whch ncludes all possble values of target type n the sea battlefeld All elements n Θ are mutually exclusve So Θ s a recognton framework of the sea battlefeld s target type Defne a map m : Θ [0,], t satsfes: m( φ) = 0 (7) m( A) = Θ A 73

6 Where m s the basc belef assgnment of Θ, m (A) s the value of basc belef assgnment of A When A Θ, m (A) represents the precse level of the proposton A s trust; When A = Θ, m (Θ) s the degree of uncertanty Ths fuson algorthm s focus on the targets n the sea battlefeld, so the target fuson recognton frame s defned as a set of warshps Surface targets are manly destroyers, crusers, frgates and arcraft carrers n the sea battlefeld In the entre arcraft carrer battle group, destroyers and crusers have the smlar sze, poston of the fleet and performs the smlar functons Therefore, destroyers and crusers are classfed as a class n the mult-feature fuson recognton framework The framework of target recognton contans three elements (destroyer or cruser, frgate, arcraft carrer), so Θ can be expressed as a set of Θ = { D, F, C} In the set, D represents destroyers and crusers, F represents frgates and C represents the arcraft carrers sng the basc belef assgnment functon to descrbe the uncertantes n the process of target recognton, we defne a basc belef assgnment functon m whch meet the followng condtons: k m k ( φ) = 0 (8) m k ( D) + mk ( F) + mk ( C) + mk ( Θ) = (9) Where k s the number of target recognton features; m k (D), m k (F) and m k (C) are the value of basc belef assgnment of the dfferent knds of targets n the mage, whch are computed by mage features; m (Θ) represents the degree of uncertanty n the decson process In order to get the basc belef assgnment value of mult-feature fused, we use dempster fuson rule to process each feature confdence value Dempster fuson rule: Let m, m,, m represents the value of basc belef assgnment n the dentfcaton framework Θ, whch comes from dfferent sources of evdence; through k orthogonal and composte the fused basc level of confdence can be gotten [ m j ( A )] A = A j k m( A) = ( m m m j )( A) = (0) [ m ( A )] Where k A = φ j k A A, Θ ; m s the basc belef assgnment functon whch s constructed by mult-feature evdence Accordng to the calculaton result on above, the followng two formulas are the Bel(belef functon) and the Pls(plausble functon) Bel ( A) A ) () A A A A φ Pls( A) A ) () j 74

7 Where A, A Θ ; Bel (A) denotes A whch s really trust; Pls (A) denotes A that s really not negate degrees For A Θ, A s confdence level equal to the sum of the value of basc belef assgnment whch s computed by all subsets n A From the above, target type verdct probablty nterval s [ Bel ( A), Pls( A)] In the fuson recognton framework, dfferent target s Bel and Pls can be expressed as: Bel( D) D) D : (3) Pls( D) D) + m( Θ) Where D probablty nterval s [ Bel ( D), Pls( D)], D Θ Bel( F) F) F : (4) Pls( F) F) + m( Θ) Where F probablty nterval s [ Bel ( F), Pls( F)], F Θ Bel( C) C) C : (5) Pls( C) C) + m( Θ) WhereC probablty nterval s [ Bel ( C), Pls( C)], C Θ 4 Instance Expermental research envronment: AMD Athlon Dual Core Processor GHz, 40GB RAM, Wndow XP, Matlab 009a The expermental data comes from Google Earth, the mage resoluton s better than 4m Frst step s to detect targets n the mage and compute the target features as part descrpton The followng s to use the results of part as evdence and put all evdence nto the mult-feature fuson framework as part 3 descrpton Wth the mult-feature fuson decson, Bel and Pls are appled to calculate the confdence nterval of targets Fnal step s to get the target type nformaton and ts confdence Table 3 s the target recognton results by usng our algorthm Table 3 Target recognton results Target Num Target Image Warshp Destroy(Cruser) Frgate Arcraft carrer 39% 5% 936% 33% 669% 08% 3 696% 67% 37% 75

8 4 86% % 9% 5 356% 636% 08% From the table 3, the confdence of target and target 4 s over 85%, whch concern target s arcraft carrers It s determned to be targets of arcraft carrer The confdence of target and target 5 s over 60%, whch concern target s frgates, so targets of frgate can be found The confdence of target 3 s over 65%, whch concern target s destroys or crusers, t can be determned targets of destroy or cruser Comparng the expermental results wth the mage data, our algorthm can correctly dentfy shp targets 5 Conclusons Ths paper proposes a shp target recognton algorthm whch s sutable for sea battlefeld The research extracts the feature nformaton from target mage and use mult-feature of the shp to establsh the mult-feature evdence fuson frame based on D-S evdence theory Because sngle feature s dffcult to recognze the target and make a wrong decson possbly, By the mult-feature evdence fuson we can mprove the ablty to recognze the target The expermental results show that our algorthm can relably and accurately dentfy shp targets n the sea battlefeld and mprove the recognton ablty of the system References Zabd M M A, Mustapa I, et al Embedded vson systems for shn recognton [C] IEEE TCENCON Washngton DC, SA: leee Computer Socety, 009:-5 Luo Q, Khoshgoftaar T M, et al Classfcaton of shps n survellance vdeo [C] IEEE Internatonal Conference Informaton Reuse and Integraton Washngton DC, SA: IEEE Computer Socety, 006: Luo Q, Khoshgoftaar T M, et al Classfcaton of shps n survellance vdeo [C] IEEE Internatonal Conference Informaton Reuse and Integraton Washngton DC, SA: IEEE Computer Socety, 006: Zhang Z A Study on Harbor target Recognton n Hgh Resoluton Optcal Remote Sensng Image [D] Hefe: nversty of Scence and Technology of Chna, Lan J, Wan L Automatc shp target classfcaton based on aeral mages [C] Proceedngs of SPIE Bellngham Wash: SPIE, 009, 756(): -0 6 A P Dempster pper and Lower Probabltes Induced by a Mult-valued Mappng[J] Ann Math Statst Volume 38, Number (967),