Զѧ
AGTA AUTOMATICA SINICA
1999 25 2 Vol.25 No.2 1999



δ֪Դλû״ָ緽о
Ӣ̵¡״
ժҪ緽δ֪Դλûά״ָ.棬緽֪״弰ӦͼĻҶֵѧϰȨֵΪԴ.ɴ˿ɻָͬԴά״.ʵ֤緽Խδ֪Դλû(Դ)ά״ָ.
ؼʡ磬״ָ.
SHAPE RECOVERY IN UNKNOWN ENVIRONMENT 
BY NEURAL NETWORKS
TIAN YingliMA Songde
(National Laboratory of Pattern Recognition
Institute of AutomationChinese Academy of Sciences, Beijing 100080)
H.T.Tsui
(Dept.of Electronic Engineering,the Chinese University of HongKong)
AbstractIn this paper, we propose a new method based on neural networks to recover the 3D shape of an object with Lambertian surface. The method can be applied under multiple light sources of unknown distances and sizes. A multi-layer neural network is used to learn the mapping between the image intensities and the corresponding surface normals. A sphere is used as the calibration object in training the neural network by a backpropagation algorithm. The weights of the network can be regarded as the environment parameters. The efficiency of the method is verified by simulated and real experiments.
Key wordsNeural network, shape recovery, Lambertian surface.
1
״ָ⣬˸ָĽɻҶȻָ״(shape-from-shading)巨(photometric stereo).ҲڼչԴµ״ָ.ЩйԴλû֪.ܽδ֪Դʵʻе״ָ⣬ɿ԰ά״ָоʵӦƽһ.õ㷺Ӧ.Lehkyˣ1״ΰɻҶȻָ״(shape-from-shading).öǰͨǧ֪״ͼѧϰȻõδ֪״.Iwahoriˣ23緽ɾ淴ָ״巨(photometric stereo)еı.Що緽ЧӦ״ָ.
һֻһͼָδ֪Դ״·.ûֱӼı״öǰƻĹԴ.һֻֻһ֪״ѧϰ(ͼҶֵ)(淨)֮Ĺϵ.ΪͼӰ켰Լÿһ33ڵľŸصĻҶֵӦص淨.ʵ֤緽Խδ֪Դλûά״ָ.
2״
2.1ʵʻά״ָ
ʵʻеĹԴ󲿷չԴιԴԲιԴǵϵĹ.ϵPչԴιԴԲ͹Դ£B˵㷨߷n(n,n)ĹϵΪ

ΪԴiΪԴ.
ԴA,B,C֪״ָõ״.緽ָδ֪Դλûά״.
2.2ǰά״
ʵʻеĹԴδ֪.緽ԹԴй.һ֪״(ĳΪУ׼)ķʸΪδ֪ԴеͼΪѵӶõʾȨֵ.Ȼ󣬱ôָͬı״.Ӧǰ磬Ԫnx,ny,nzʾ巨ʸn



NΪڵwj1(2),wj2(2),wj3(2)Ϊڵ㵽ڵȨֵw1(2),w2(2),w3(2)Ϊڵ123ֵ(x)=1/(1+e-x)Ϊsigmoidfj(u,v)Ϊjڵ룬

Ii,i=1,2,,9Ϊͼ(u,v)Ϊĵ33ڵĻҶֵwij(1)Ϊiڵ㵽jڵȨֵwj(1)Ϊjڵֵ.
һΪУ׼ѵ.ȨֵС˽ͨб淨EСֵã


ʽDnΪͼnx(u,v),ny(u,v),nz(u,v)ΪУ׼϶Ӧص(u,v)ķ߷ʵֵx(u,v),y(u,v),z(u,v)Ϊô˵㷨߷.ΪʹʽEСBP㷨.ԴӽڵkڵlĵmѵĹΪ


ʽЦΪѧϰnetlΪڵl룻okΪڵk.ڵlΪڵ㣬Ϊol(o1=x,o2=y,o3=z).

ڵlΪڵ㣬

3緽ָδ֪Դλûά״Ӧ
Ӧ(㡢)ʵ״ָ.ڵΪ112.ͼߴΪ128128淨߳ߴΪ126126.ΪͬʱӦͼҶϢԼÿһ33ڵľŸصĻҶֵΪ룬Ӧ˴ڵص淨ʸΪѵ.봰ڵƶΪһ.
Ϊ˵õȷȨֵйһ.ͼҳͼҶֵȻʹÿһصĻҶֵͼҶֵ.һͼĻҶֵ01֮.

ı淨nx,ny,nznx,nyȡֵΧΪ-1, 1ݣnzȡֵΧΪ0,1.ʽnx,nyȡֵΧһ0,1ݣ

4ʵ
4.1ģͼʵ
ù׷ٷڵԴͶԴ·ֱͬɫͲͬʵĲ֤ͬ˷.һ֪״У׼ͼѵ.ʵδ֪Դλû(Դ)壬緽ЧĻָԴµĲͬɫͬСͲͬ淴ʵ״.ƪƣֻʵͼʵ.
4.2ʵͼʵ
ʵУιԴһԴ.һ16mmͷCCDȡͼ.ʵͼ1ʾ.(a),(b)ʵ.Կ緽δ֪ԴʵҲԺܺõػָ״.


ͼ1緽ָĶԴʵʻͼ
5ۺ
һЧ緽ָδ֪Դλû(Դ)ά״ķ.棬֪״弰ӦͼĻҶֵѧϰȨֵΪԴ.Ȼָ˹Դеά״.ʵ֤Чػָı״.ԱĹһо1)м״ΪУ׼ѵ磬Ը״ָǷ㹻2)ǵʵУ33봰ʺϵ.Ըӵ壬봰ڴСӰһо.
߼飺ӢŮ 1966.1996Ĵѧʿѧλ.йѧԺԶоģʽʶصʵҹ.оΪӾͼѧͼ.
̵¡1946Ϻ.1968ҵ廪ѧԶϵ1983ѧʿѧλ1986ѧScience d'Etatʿѧλ.19831984ΪUniversity of MarylandĿоԱ19841986ΪINRIARobot Vision GroupĿоԱΪйѧԺԶо.ҪоΪӾ񾭼ͻڴĻ˿. 
ߵλӢ̵¡йѧԺԶоģʽʶصʵҡ100080
״Ĵѧӹϵ
ο
1Lehky S RSejnowski T J. Network model of shape-from-shading neural function arises from both receptive and projective fields.NATURE.,1988, 333-2452454
2Iwahori Y, Tanaka H, Woodham RIshii N. Photometric stereo for specular surface shape based on neural network. IEICE Trans. INF. & SYST.,1994,E77-D(4)498506
3Iwahori Y, Woodham R Bagheri A. Principal Components Analysis and Neural Network of Implementation Photometric Stereo. Inproc. of the Workshop on Physics-based Modeling in Computer Vision,1995 117125
ոڡ1997-08-25
޸ĸڡ1998-07-15
