о뷢չ
JOURNAL OF COMPUTER RESEARCH
AND DEVELOPMENT
1999 36 12 Vol.36 No.12 1999



Ĳλģͼʶо
¸塡꿵
ժҪԶʶһѵҪĹ.һֻλʶ𷽷.÷ȶп׼ȷλ׼ȻԪֱԶλȡӶõڵͷֱʺͽϸ߷ֱϵʶ𣬾ʶʸߡСڴʶص.˷1300Ͻ˲ԣת仯δѵԿԺܺõʶ.
ؼʡ׼ԪλͶӰ
ͼšTP391.4
THE LAYERED FACE REPRESENTATION MODEL AND
FACE RECOGNITION
ZHANG Gao-Qing and WANG Shen-Kang
(Department of Computer Science, Zhejiang University, Hangzhou 310027)
AbstractAutomatic human face recognition is a difficult but significant problem. A novel method for face recognition based on layered face representation is put forward in this paper. The method first locates the face and key facial features including eyes, nose, and mouth in image quickly, and then normalizes that face depending on the center location of eyes. Then PCA neural network is used to extract characteristics of the located faces and facial features. This can be viewed as a layered representation of faces: where a coarse and low-resolution description of the whole head is augmented by additional high-resolution details in terms of salient facial features. This method has been tested on 1300 facial images and shows better performance than traditional PCA for face recognition, especially when head poses and expressions are changed.
Key wordsface normalization, principal components analysis (PCA), layered representation, integral projection
1
һ޶ʹһ˫̥ҲĳĲ.Ȼڱ顢͵ȷ޴仯ԿɺѵؼʶҪһܹȫԶʶϵͳȴǷǳѵģ漰¼⣺ٰһģʽΪۻķ.ܹʵܵϵͳй㷺ӦǰɷƬʶ𡢶ʧͯϢȫϵͳÿȷϵȣڹȥ20УԶʶڶоߣĿǰȫЧϵͳδƳ.
еʶ𷽷ɷΪ12ݣһǷġھֲʶġȫʶ.ǰ߷ĸɲ֣ȡɢľֲΪȻΪñ׼ģʽʶʶ.ϲ׽Ϊһȫֵ.ЩֻעؾֲĳһһҪ仯̫ת仯ʱܴﵽЧ.⣬ڲԪĻϣ෽нϣĲλģʶ𣬴Ӷһ仯ʱϵͳʶ׼ȷ.
ڶп׼ȷλͱ׼֮󣬱ĲԪPCAԶλֱȡõӦȫ;ֲ㣬ͷֱĴֲͽϸ߷ֱͻϸ͹Ĳλģ.ڲλģ͵ʶ𷽷˷˵PCA㷨34ݵһЩȱ㣬ٻᱻת仯ƭڶûѵʶϲ.ͬʱPCA㷨гȡٵص㣬ڱʾλģ͵УӦʶ.ļٶͼֻһұΪ.
2׼λ
׼ֲĶλʶǰͻֱӰʶЧ.ˣ׼ȷı׼λʮֱҪ.
2.1׼
׼ͬʱҲ۾Զλ˼ǻһ򵥵ʵеͼͷ۾ԷֻֿҪͼܱܺõطָͿɽ۾׼ȷͬʱĶλ׼ɵõ׽.׼ͼСΪ128192أͬһˮƽλ.
2.1.1ֵָ
ͼI(x,y)ֵָʹͷ۾Էֿ.ڼٶͼֻһָԼ.ͼ1bͼĵֱͼԵķֱֵӦİͷ۾üëȣͷյȣ.ҵһֵӦӦΪһ㱳ܱİֻҵǰֵȡһҶֵΪֵ.ڴڶѡֵͨԶֵʹͼĶ̷ֵǳɿ.
ֱͼгͼĻҶȷֲŵӰһЩƷŵ㣬ǵĴڻᵼ¶ԷУӶӰֵѡȡ.Զ֮ǰֱͼƶƽƽԤ

ʽGiʾҶֵΪjƶƽֵgiʾֱͼлҶֵΪiسֵƵnΪÿƶƽֵõĻҶȼ.ƶƽʵǶɢмƽŶ.һƶƽԲҪɶGinȡֵͬλƶƽ.һжƶƽɴﵽЧ.ͼ1cǶͼ1bƶƽĽ.ɼԭֱͼԸı䣬ŵ㱻.
ԣͨԭʼͼֱͼƶƽ׼ȷǰֵõɿֵзָʹͷ۾Էֿ.ͼ1dֵָĶֵͼB(x,y).

ͼ1ԭʼͼֵָ
2.1.2۾λ
һ۾λģƥ䷽12.һģƥӦСͬYilleˣ5Xieˣ6ݸĽĻڵģƥȡȻȶɿԴżӦһȱ.һֳ֮Ϊ۾λ˷ֵָĶֵͼB(x,y)һЩ֪ʶԿٶ׼ȷضλ۾ңͬʱλͼ2.۾λҪ裺
(1)ͨһɨֵͼB(x,y)λͷ
ɨB(x,y)Ĺаհͨԭ򣬲ڵ㷨ǳ1ֵ.һ˵ڱΪ򵥵£һ鼴Ϊͷ飬ͨǳģʹԹͷ׷˵ͷҶҲǱȽСģԶֵҲܽͷҳ.
(2)ͨͷĴλͷڱ߽磬ͼI(x,y)
ڶλͷһδϵµľֲɨ裬Եõͷ±߽ߣͷпܳ0ֵߵ㣬ƽֵ۲ߣǷ߸һСֵ޽ǵĲλ˵ͷ쵽·.ҳ㣬һͼI(x,y)ͼΪͷ±߽ߵߵ㣬ΪҵСֵλã׶Ϊ˼Ҷ˼1.5ͼϷֽĽСֵ.
(3)ͼI(x,y)Ӧ˹ӣڶɨ裬λ˫
ʹI(x,y)еÿҶȳ¾

-1-1-1
-18-1
-1-1-1

ͼ2εĶ΢ģ

ͼ3۾λʾ
ͼǱ߽еģֵָͬȻȫͼɨ裬붨λͷͬ˴ҳ10飬ֶԴǳΪ˫۵ĳɶԿ飬бпܳΪ۾һԣȻȡһԿڵĵΪ۾ĵ.
б۾õҪ
(1)۾(width)߶ȣheight(area)Ӧ㣺
width > height and area > 0.5
(2)۾I(x,y)ϱ߽ľd1䵽I(x,y)±߽d2Ӧ㣺
d1<d2(۾Ӧϰ벿)
2.1.3׼
۾ĵ̬б׼ͼI(x,y)ĵlתlﵽˮƽλãͼI(x,y)ȻI(x,y)ű任ʹͼInorm(x,y)ߴΪ128192أСı׼.һͼInorm(x,y)Ķüëϱ߽磬׶´䣨˶죩˾ȧǸͼ4ʹؼ͹ֳͼʶķΧС.

ͼ4ͼ׼ʾ
2.2λ
۾λ󣬱ӡֲͨڻͶӰ7ݵķ㶨λ.
۾ľλãҲ֮ȷӺǡϣ۾·Ϊ׼ʱͷбѾ˫Ϊ׼ת迼ͷת׼Ӻ͵Ĵֱλÿɳȷ.
ӵľȷλÿͨѰҴֱݶͼˮƽͶӰķֵȷλԭͼˮƽͶӰĹֵȷ촽֮ڵĲ֣.ֵ͹ֵǵͻֵۣļΪӺ͵Ĵֱλ.þȷĴֱλúɽһڸСĴ.
ӵˮƽµڽУˮƽݶͼĴֱͶӰѰҸƽֵķֵҵķֱֵΪӵұ߽.ĸ߶ȿͬڴֱݶͼϵõˮƽλͨƽֵָˮƽݶͼĴֱͶӰȻѰ㼴ȷλͼ5ʾ.õֲλú󣬾Ϳɽʶ.

ͼ5λʾ
3ʶ
ʶԪPCA.ͼԿһʸͼĸ߶ȺͿȷֱΪhwӦʸάΪwh.ʸһռ䣬Ϊͼռ䣬άΪwhصͼ.Ϊƣʸۼͼռһխ.ԣͼռ䲻һŻռ.Ԫǹһܸռ䣬ͿռάʹµռĻΪԪܸõ͵ģʽ.
3.1Ԫ
ԪȻͳѧչģ˹¶岢õ㷺Ӧ.Widrow-Hoffؾ洢LAMM8ݿɿһԪ磬ĿǲһӽѴ洢ķӦڴ˼Ϊһ.LAMMһ㹹ɣÿԪʸеһԪӦÿwhԪÿԪԪͼ6ʾ.

ͼ6ؼĽṹ
LAMMѵвWidrow-Hoff򲻶ϸıȨֵʹﵽС
wij(t+1)=wt+ǡ(xj-oj)xi
(1)
УxiΪʸĵiԪأojΪjԪwijΪijԪ֮ȨֵΪѧϰ.LAAMԣԪojxjȨֵĽԤڲͬı.
I=whΪظΪʸάKѵX=xijάΪIKѵеkxkӦѵеkұһW=wijΪάIIȨֵ.ôѧϰ1ֿдΪ
W(t+1)=W(t)+ǡ(X-W(t)X)XT
(2)
ʽ2˵˹.
XдΪX=PQT.УPΪXXTQΪXTXǰXXTֵĶԽǾ.ɶ壬PPT=IQQT=IIΪλʽ2ɵã

ԣʽ2дΪ
W(t)=P(I-(I-(ǡ)t))PT
ȡΪXXTֵԣѧϰ̽ʱȨֵΪΪ
ֿɿy=PTxo=PyɣxΪһʸoΪyмʽPʸxͼռתռеʸyPʸռĻԪ.ʾмʽͼ7ʾУrank(XXT)=min(wh,K)ͨKСwh¹ɵռάͼռά.
