Recognizing human faces

Recognizing human faces

A ppfied Ergonomics 1975, 6.2, 104-109 Recognizing human f es G.H. Fisher and R.L. Cox Department of Psychology, University of Newcastle-upon-Tyne I...

610KB Sizes 0 Downloads 65 Views

A ppfied Ergonomics 1975, 6.2, 104-109

Recognizing human f es G.H. Fisher and R.L. Cox Department of Psychology, University of Newcastle-upon-Tyne

Information as to identity appears to be concentrated in and around certain areas of shapes and objects whereas other parts are more-or-less redundant. The experiments described in this paper were intended to establish which particular features of human faces, both in isolation and combination, convey most information for recognition. The resu Its indicate the relative importance to recognition of different facial features. However, certain faces appear to embody idiosyncratic cues while some are confused consistently with others whom they appear to resemble.

Most people believe themselves to possess a remarkable ability to recognize others, particularly their faces. The apparent self-evidence of this ability is such that it is seldom called into question. However. the fallibility of human recognition has recently been demonstrated very clearly by failure of observers to reconstruct adequate representations of criminals, or suspects, with the "identikit" and "photo-fit' systems employed widely for these purposes by the police authorities of North America and Great Britain respectively. Recognizing a particular stimulus for what it is. or represents, requires first, that it is allocated to membership of a particular category of stimuli and. secondly, that it is identified uniquely within that category. Bearing m mind the sizes of populations from which stimuli might be drawn for experimental purposes, therefore, recognizing them correctly must demand extremely precise classification by the perceptual mechanisms concerned. Recognition of twodimensional shapes drawn from small sub-sets of patterns has been studied extensively. In general, it appears that certain parts of such stimuli contain considerable information whereas others include relatively little, or are altogether redundant. Hence. the former convey most of the cues required for recognition whereas the latter contribute few, or none. However, comparatively little is known of any differential cue-content of more commonplace objects. Remarkably few studies have examined the facility with which familiar stimuli drawn from very large populations of such objects may be identified. Most people find themselves able to recognize the faces of numerous acquaintances, friends, colleagues and public figures al first sight. Thus, taken as a whole, the human face must be particularly rich in information. Nevertheless, il seems reasonable to assume that different facial features vary in the amounts of information they convey. The intention of this paper is to determine whether any differential contribution is made to recognition by the several component features of human faces. In the most systematic previous attempt to study facial recognition, Goldstein and Mackenberg I1966)"explored what features of human laces carry information" about

104

Applied Ergonomics June 1975

identity. They presented 76 children with photographs of their class-mates, masked so that only certain features were exposed, and found "that various parts of faces differ in how much they contribute to recognition". The upper and right-hand halves of the faces were identified equally well. Therefore, Goldstein and Mackenbergconctuded that "'upper portions of the face are more helpful in identification than lower portions". Upper right quadrants of the faces. however, were identified only slightly more readily than those revealing hair and brow detail only. This suggests thai the eyes themselves convey relatively little information tot recognition and this was confirmed by results obtained when presenting either one eye, or both, in isolation. Tile mouth and other lower parts of the lace alone were also found to make relatively small contributions to recognition. This indicates that more cues are embodied within combinations of features than individual parts. Goldstein and Mackenberg noted further that "'Those parts of faces which have high lot low) recognizability for young children also have high (or low) values for older children". But they also found thatthe ability to identity faces was correlated significantly with age. Those findings were confirmed in a later study made by Chance, Goldstein and Schicht ( 1967 ) who made the further observation thai children were able to recognize faces of their friends somewhat more readily than others. These previous studies were confined to American kindergarten and grade-school children. Hitherto. no systematic investigation has been made of facial recognition with relatively large and representative samples of adult subjects. Such an investigation was attempted by the present experiments.

Experiment 1 Careful consideration was given to two important methodological problems before tiffs study was embarked upon. First. results of experiments ill which subjects are required to recognise stimuli in some appropriate way may be confounded by any differential familiaril y of materials used. As Goldstein and Mackenberg acknowledged, one weakness of their study was "that older children may have

Table 1: Frequencies with which male and female subjects portrayed in samples of 1000 photographs posed so as to face towards left, right, or straight-ahead with respect to observer Left

Straight

Right

ahead

Male

502

197

301

Female

363

111

526

known each other longer than younger children, especially in stable school populations". Their insistence that "the kindergarten subjects had been together for at least seven months prior to the testing session" fails to overcome difficulties arising from different lengths of personal acquaintance, particularly with such small sub-samples of young and socially immature subjects {five groups each comprising 15, or 16, children only). The relevance of this problem was emphasized by Chance et al's further study which, as noted above, revealed the influence of personal friendship upon recognition. Furthermore, both these previous studies overlooked the possibility of a substantial contribution having been made to their results by the number of occasions upon which participant subjects were given opportunities to identity features of their own faces. Secondly, photographs of component parts of human faces may be arranged in innumerable different ways. Goldstein and Mackenberg, along with Chance et al, decided arbitrarily upon 13 divisions and displayed them to subjects in randomly assigned order. Close inspection of their results reveals that the contribution made by component features to identification fail to predict those expected from comparable composite arrangements. Hence, from these earlier studies, it is impossible to establish whether incremental increases of information progressively enhance recognition in any systematic way. Materials The Picture Library of Thompson House in Newcastleupon-Tyne contains more than five million photographs. Over half these photographs feature people, many of whom are well-known both nationally and internationally. Having been acquainted with the purpose of this study, the Librarian kindly agreed to have his staff make a selection of full-face photographs of those whom they considered

"would be recognised Cbirly readily by average adults". This proved considerably more difficult than it might appear at first sight since full-face photographs appear relatively infrequently in press and agency collections, and celebrities in particular have a tendency to adopt poses presenting their faces in three-quarter profile: males looking towards the left and females towards the right. This was an unexpected finding which the authors were acquainted with informally by the Library staff who made the search for photographs. We asked them to draw a random sample of 1000 photographs featuring one person only so as to enable this tendency to be verified. Table 1 shows how direction of pose is related to sex of portrayed subject. The X2 value referring to this contingency table is 406"5 ( p < 0"001 ). The immediate implication of this result appears to be that celebrities tend to pose so as to present the partings of their hair towards the photographer and thus expose more of

their faces themselves. Nevertheless, some 67 photographs were eventually selected by the library staff. Unfortunately, many of these were immediately found unsuitable for the required purposes because the faces were partially obscured in some way; by other people, adjacent objects, hands, hats and various articles of dress. Further photographs contained such idiosyncratic cues to recognition as pipes and spectacles. All of these, along with any including growths of facial hair, were discarded. The only remaining photographs portrayed the following people: 1. 2. 3. 4. 5. 6.

HRtt Princess Alexandra Dora Bryan Violet Carson Barbara Castle Henry Cooper Robert Dougall

7. 8. 9. 10. 11. 12.

Lord Hailsham Margaret Lockwood Lulu President Nixon Cliff Richard Emanuel Shinwell

Readers may express reservations as to whether members of the general populace might be expected to recognise even studio portraits of all these people and the photographs used in the present experiments fell somewhat short of such quality. The authors certainly had such reservations at this stage of the inquiry but these were virtually the only materials found suitable for the intended purposes in one of the largest collections of photographs available. The original photographs differed appreciably in size and several copies were required of each. They were all rephotographed and the faces only reproduced m standard size as glossy prints; 4 ~ n (115 mm) long and 3m (75 mm) wide. From these prints four series of materials were prepared. The first two series were obtained by cutting five photographs horizontally across the following levels: (i) immediately below the mouth; (ii) between the mouth and nose; (iii) across the top of the nose so as to include the nostrils; (iv) immediately below tire eyes and (v) immediately above a line defined by the eye-brows. By adding one complete photograph to each of the two sets of pieces so obtained an 'Up-Down' and a 'Down-Up' series of part-to-whole t:aces were prepared. Examples of these materials are illustrated in the upper two rows of Fig. 1. The second two series of faces were obtained following a similar procedure in which five further photographs were cut vertically downwards: (i)between the left eye and ear; (ii) between the left eye and nose; (iii) down the centre line of the face exactly bisecting the nose; (iv) between the nose and right eye and (v)between the right eye and ear. These are referred to henceforth as 'Left-Right' and 'Right-Left' series and examples are shown in the two central rows of Fig.l. Procedure The four series of laces were mounted separately m 12 six-page booklets in appropriate part-to-whole order. The booklets were arranged in a different random order each time the study was carried out. Four hundred psychologically naive subjects aged between 15 and 77 years (X = 26'56: 6X = 1.14 years.), all of whom were residents of Newcastle-upon-Tyne, participated. Each subject was presented with one only of the four series of photographs. All subjects were given the following standard instructions verbally: "I have here some photographs of parts of the faces of several well-known people. Would you kindly look

Applied Ergonomics June 1975

105

(*)

q II (iii)

(iv)

(v)

(vi)

Fig.1 Examples of the experimental materials prepared in different part-whole orders

106

Applied Ergonomics June 1975

at each of them carefully and try to identify the people by name. If you think you know who they are but cannot remember their names, please try to tell me what they do, or what they are well-known for".

I00

80 Lo

Results Subjects were deemed to have recognised the faces either if they were able to identify them correctly by aame, or if they indicated with reasonable precision what the particular person portrayed was most well-known for. Thus, such responses as :"She used to be our Minister of Transport", or: "He reads the news on television" were accepted as identifying the subjects of photographs in No 4 and 6 series respectively. The results of this study are plotted graphically in Fig.2. Stages of presentation are referred to on the abscissae. The ordinates indicate the percentages of those subjects who finally suceeeded in identifying the photographs, of recognising them at each level of exposure. Thus, the extreme right-hand points of all four polygons have been placed arbitrarily at 100%. Naturally, the ability to recognise faces is expected to increase as more features are exposed. This tendency is evident generally throughout all these results. Noteworthy individual differences between responses made to particular people and series of photographs will be commented upon in discussion. Fig. 1a shows that upper features of human faces convey appreciably more information for recognition than lower features. Differences between results referring to these two series are statistically significant throughout the entire range. In the 'Up-Down' series, the eyes make more incremental contributions to recognition, from 11.53% to 63.89%, than any other individual feature. They do so likewise in the 'Up-Down' series, adding from 64"40% to 89-54%. Thus, both series reveal the relative importance of eye-features in the different contexts. Nevertheless, the 'Up-Down' series indicates that lower features of the face also convey considerable information since, even in the absence of eyes, 64"40% of the photographs were recognized correctly by stage 4. In the 'Left-Right' series, 13"44% of the photographs were recognized from extreme left features alone, corresponding with 7"27% for extreme right features in the 'Right-Left' series. (Note that the left-hand parts of photographs represent right-hand sides of actual faces and vice-versa. ) This is the only statistically significant difference between the two sets of results plotted in Fig.2b. The second items in these series make the largest incremental contributions to recognition of all - up to 79.04% and 78"42% respectively. This suggests that, in this context at least, exposure of one eye only serves equally well as both. Moreover, differences between these latter values and those referring to Stage 2 of the 'Up-Down' series indicate that the additional lower parts of the face contribute 14-5% and 15.15% respectively more information when present within the contexts of upper features.

Experiment 2 The results of the experiment described above indicate that both upper and lower features convey appreciable amounts of information as to the identity of human faces. Cues to recognition appear to be concentrated in, or around, eye and mouth regions. The second study was intended to establish the relative importance of these particular features and surrounding regions.

I ~op-

90

8

up- do.n

70 60 50 4C 5C

//~

~

20 I0

1 o

I

I 2

I ~

4

I 5

I 6

Fig. 2a Stages of recognizing "Up-Down" and "Down-Up" series of photographs I I00 9O 8O tn

60 ,Rig,ht- left

5O

g c_)

~

4o 3o 2o I0

b

I

2

3

4

5

6

Fig. 2b Stages of recognizing "Right-Left" and "Left-Right" series of photographs

Materials Additional prints of the 12 photographs used for Experiment 1 were prepared in two further part-to-whole series. The first began by displaying the eyes only. To these were added noses, then mouths, jaws, the two sides of the faces together, and finally brows and upper hair-lines. The second series started with mouths only, proceeding by adding noses, eyes, side-face details, jaws and finally brows and upper hair-lines. They were referred to as 'Eyes-Out' and 'Mouth-Out' series respectively and examples are reproduced in the lower two rows of Fig. 1.

Procedure As before, these two series of materials were mounted separately in 12 six-page booklets and arranged in appropriate part-to-whole orders. Two hundred subjects whose ages ranged from 17 to 58 years (,X = 26-82 : 6~ = 1"16 years) participated. Half the subjects were allocated at random to one of the two parts of the study. Naturally, it was necessary to exclude from the second experiment any subjects who had taken part in the first. All procedures adopted in this experiment were identical with the last. Results The results of this experiment are displayed graphically in Fig. 3 adopting the same conventions as before.

Applied Ergonomics June 1975

107

Tire previous features

experiments

showed

t h a i . el all i n d i v i d m ~ l

o f tire h U l l l a t l f a c e . t i l e e y e s Lilt_' Iili)',l i l l l p o r l a l l r

7~

providing cities to recognition, t l o w e v c i , w h e i i eye-det:li': alone was revealed in the 'Eyes-Out' series, the\ welt' ~ecognised upon only 12.0007 occasitms, t{xperiineili :ii,~ showed

that lower

regions of the face. ii~cluding

make substantial contribtitioits t o r e c o g ! n t l t ) i l . little evidence for any high i n f o r m a t i o n a l - c ~ u i t e n mouth

in F i g , 3 s i n c e tiffs [ e a t u l e

ili i s o i a t i t m

+!

inoulll-, But

t

there

]

o I ihe

,:,s

wa> re<~gli{:4L, U

.1

Llpon o n l y 2"95(,} occasitlns, tlo\vever, Ul b o t h series. additiori of rloses added appreciably t~ recognition .......... in the "Eyes-Out' series and 15.085; in die "Mouth-Oui' series. Thus, eyes and noses combined pr~wide significaIH t~ more information than nrouths and noses. Neverlhele>,, even at this stage the results ,l!" both series still fall well below what is regarded in more c~mventional psycht~physl<~ studies as a 'threshold' value lie 505:: correct recogniti;m ! As checks for reliability of results, the next ,,rages <~: developing these series of materials were made the sau~c Thus, rriouths were added to 'Eyes-Out' series and eye-, t o 'Mouth-Out' series. At this stage the> were ~ecogniscd upon 56" 169+* and 58-08% respectively. Sia tistically, ~h i~ difference is insignificant. Sided:aces, including ears, we>,, added to the t\)rmer series wlfilst chins were included a* the next stage of the lattei. They contributed b} 3 4 - 3 1 and l 2-32'}; to ~ecognition respectively. This suggest-. !i~a; the importance of eye-detail is enhanced considerabl'~ h\ the immediately surrounding conlexl. All fifth stages wv~ c again identical: chins being added t~ 'Eyes-Out' and >talcfaces to
~/" /l .J

4

Y"

~d

b iS. 3

- li

":-" O L ' L ' : I M ~ ! P , >

i~th

;'(mr

L'MtI[', i/]ak ],.'[:l',.iV~

f,.t

~lllZed

i~;

{:'v,::

J~d \'h:t;:.h-.)t:

[~t~;i~

\Vhlc!~

lllic

"-t

]!]~)[k 1~1

L'ilal4C{t?li_-llv I

sJ

1 ill?,

i;

, , A ', t ~

¢

~

:

r -

-

~>

- ~

c

>

:'

+

c

Results of the studies described above indicate the relative overall importance for recognition of different! facial features both in isolation and combinalion. A5 expected, well-known people are recognized more ~eadii'. as more of their facial features a~e exposed and cerlam additions make greater c(mtributions than others, iloweve~ these summaries obscure three features of the presem findings which merit further cotnllletH. First, in several cases, single ~solated leattHes pr~,vcq disproportionately suffieieut in charactelizmg the whvk: face. President Nixon was recognized i~>m Ictt sidc-lac<. detail only by 33 subjects, for example; lulu was ideuitli:'~

Ergonomics

June ! ~ 7 5

who

- \laJcohll

\vt're

it,

!k,iiclcll,u

n]

,,..{ .pt~

\Ccoltnll~

,'

:~i. r p t l h L . ~ .

{ile..i.

,cc\vorii1%

,,ircctl\

~.,~

"lid

l/ave

,1,

~:~i;,, , --..

, ]>~i)'"rl

"

_ : 3 1 ; {i . ~ t , .

--.

.,!IC~.:

'~
-~i}otJit{

. .

..-t~t,~ "

-.

.-.

,

h ,

,,.:~ L-,,;:.~,,_ : h

!i]tlSc p i O l . l ; i \ c ' t l I'~t t h e plH;i~'L-' .iPe. •.\pCIllllel/-{.

~+, - ]

>...

.,~,;

:'

'
,,',,,:7

"

,t ! \ - ~k I ! :

h,

t,\tjcili%t...

, - ~

,h_t~

t-}H7

: L ,ca -

I'L'51]eL ! 'i,:'!

!iLl!lilt(lilt

,

;

, ..~-

\lugec!;.~<

l)~lll/ed

?ll.'SCI]{:lllOi

i]Itsl

hair-lil!u,.te!_u.

.-ppe~

,i%

;,;.."'++7"<.++!

' '"

,IO~,7 ill ih,: _,~>u ,l! t l l l l l c e s ,. , ' ~ . . . .

( 1902, p 1 4 7 ) p o i n i e d .'Jut. ' l ~ , e c o g n i t i o f i ~,ce~Y.: so simple, direct, immediate. The skilllul pl ocessing {,! ilfformation that must be involved i,~ Not itself appare,v, for the object seems familiar as soon as die eyes encoumc it". By denuding spatial patterns in appropriate wa}~, however, any particular feature conveying significantl> m<,c or less, information for recognition may be ascertained In inore conventional studies th_is demands specifying the sizes of sub-set from which particula~ patterns are d r a ~ L Procedurally, it also requires that units vf information a~, added progressively so as to enhance probability o! recognition. As Attneave (Ig541 and many others have shown, muctr information embodied will/visual dispiav, {s redundant. Therefore, with relezrence t~ recogniiicm. the rider "But a part in isolation nray he :is great as the whole" must be appended to the d d adage relating whde> with sums of parts.

"9

": .

• ,i~:

.~, . . . . . .

+-,

. . ' t '.~ ' r. . .

:~

HL'] .~":L

~

h,'t

-.

~\,[!.,

,~,l!iC't~

:

!I]OM

v

;itli~l\ .,

, : ~ I~% ~

;~

{':

"'.;iil!'d'

;

e,_!

type".

L ,'|

:~?,

!~1 ' e l 'it']

--

c -

;;die

,

-~:-:

lit' pht~toe~ r o t >

i.

P,]OS

~ct~,,- 4: I el,;

{li

tlil[Lll

.. h U " n "

Li ,~ .... • . " . -,.~, . . "

.ILL' e X . l g 2 2 t

tllltl!\

avcd

"<' %'l:_tl

.',::

ii1a]-#~

tl!!]'ili~{,l[;IF!{

'c'.tlfll{'. 7~ec

ill:

L ~]lleit)icled

iiil't~{]l:lt!

t,t.!llvel~

~tll

IOllt

,,pl~c~ h a u . l h v .

~cl)e:lledlv

As Miller

Applied

v/

St-ages o f r e c o g i u z i i ~ g

t~!l,~, l i O l l t

108

I

;eiieg of photograpll7

lhe

Discussion

;

if

,!

Lit

~-'(~ ;!a

;'

,"

,:,'-.

" , ~I ~~ i 1, i ~

--

~¢' "

k ' .I :

. .

0v l'.

l

ill

lcDalal01£

flu,let

Huickly

bee_tile evident

dillered

wide[',

:exults Dl~)!icd who

linall~

•: , , r : v ' c t / , . xtllt)SC

,_ -

i l i a : -;~.

H

t-iTs

~tv,.coeded

tlSt_'(It]

tilt: ofiClna '.lllllildlJl\

}~\

'

,,

,, .~<, ,,

~Taviw

~

,:,],,;:,.-

-

LIII\

~.',t

0LIdt~;>

tll

!

.,

,:,,,..

"

...

'~. / , :

'!,> .H."I::

, ' ~ I ; ~ [ - , 1,

..... ,,<.,

,it

,iui,:-

,:'-

....

V.|\|~

_kit:2

' ~,,

.t::J.';L,,~

~-I ,i ,._ "- ~,~lh4r]r~ '

iiC:Jll]ICil! ,~ '~:

, ~:

~ .~.,:

!]:i{:I {!ldic:illll~' 2 'l'he~

--

m tainilicl~ti . . . . . . . . .

phc)t,lgr,q)hs wcic 'l>:cl

[:ibk ]O111](1

e,Hdlt

!:ii,,i,!,,, ,,'

'

- - ; " +, I - H

tl'}h/i.-:H::'

i

Acknowledgements t h i s re.~,.,:t~i~ w a s ~ u p p t ) l ! : : Q t,,uncil

(;:au,: No tlR

(.)t.i~ i h a i l k : , dr¢ d t l c t { photoglaph: iuv-duabL:

suitable

--~ i::i ,>ci'.i ~, ~,-:.,c-:<,.

!J-2~

t~

~}i{: i : t ,.::+r;: ; ;

,: :.i;; ~ir i { , - l ; 7 {

t},.7-:: ,~d

,~ "

Table 2: Frequencies w i t h w h i c h subjects of p h o t o g r a p h s were recognized at final stage of p r e s e n t a t i o n

0r,0,

Up-Down

Down-Up

Left-Right

Right-Left

Eyes-Out

Mouth-Out

H.R.H. Princess Alexandra

65

79

71

62

49

54

Dora Bryan

78

88

86

72

85

79

Violet Carson

96

98

96

94

99

98

Barbara Castle

93

100

95

89

93

99

Henry Cooper

92

91

94

85

88

95

Robert Dougall

59

79

78

70

73

80

Lord Hailsham

72

85

80

72

57

77

Margaret Lockwood

31

58

40

33

30

42

Lulu

92

98

99

96

99

97

President Nixon

95

96

98

93

99

98

Cliff Richard

93

92

99

96

93

97

Emanuel Shinwell

45

78

61

46

52

67

References

Attneave, F. 1954 PsychologicalReview, 61,183-193. Some informational aspects of visual perception. Chance, J., Goldstein, A.G. and Schicht, W. 1967 Psychonomic Science, 7, 223-224. Effects of acquaintance and friendship on children's recognition of classmates' faces.

Goldstein, A.G. and Mackenberg, E.J. 1966 Psychonomic Science, 6, 149-150. Recognition of human faces from isolated facial features : a developmental study.

Miller, G.A. 1962 'Psychology, the Science of Mental Life' New York : Harper and Row.

Applied Ergonomics June 1975

109