Você está na página 1de 28

u

e
an art anthology

no. 75 (XXXVIII), 2017


Cracow - Vienna
Conrad Rudolph, Amit Roy-Chowdhury,
Ramya Srinivasan and Jeanette Kohl

FACES: Faces, Art, and


Computerized Evaluation Systems
-A Feasibility Study of the
Application of Face Recognition
Technology to Works of Portrait Art

Introduction

On 11 June 1685, James Scott, duke of Monmouth, first duke of


Buccleuch, and illegitimate son of Charles II, returned to Eng­
land from Amsterdam as the self-proclaimed head of the anti­
Catholic party with the intention of seizing the throne from his
uncle, James II, the last Catholic monarch of Britain. Defeated
at the battle of Sedgemoor and, two days later, dragged from
a ditch where he had been hiding, Monmouth was condemned
to death by the king and beheaded on Tower Hill in London,
the executioner requiring at least five strokes to finish the task.
A very unusual painting that may be connected with these
events is held by the National Portrait Gallery in London: a por­
trait, unquestionably, but one of a relatively young man lying in
bed with his head on a pillow, eyes closed, and blanket and
sheet pulled up under the chin, but with what seems to be anoth­
er small piece of white cloth between the sheet and the man's
throat [Fig. 1) .1 Some think that the figure represents Monmouth
-with his head put back in place, the severing perhaps wrapped
by the small piece of white cloth, lying in state, as it were. Others
think that the image does not depict Monmouth. One way or the
other, given the particular associations of the painting with Mon­
mouth and the importance of the events involved, to determine
who this person was would potentially be more than a simple
restoration of the identity of the subject. It would potentially be
Conrad R udolph, Amit R oy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

1. Artist unknown, identity of sitter in question, c. 1640s, London, National Portrait Gallery, NPG 1566. Photo: National Portrait Gallery

266
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

2. The FACES anthropometric distances 3. The FACES local features

the restoration of a unique 'document' in the history of a forma­ In the application of this technology to actual (that is, pho­
tive period of a great nation, one that, in its own way, might be tographed) human faces, a number of difficulties are inherent in
seen as a commentary of sorts on the larger events of which its a real or perceived alteration of appearance of the face through
depiction may be a part. variations in facial expression, age, facial hair, angle of pose,
lighting, and the very identity of the subject. With portraiture in
sculpture, painting, and drawing, not only do all the problems
The FACES Algorithm and Initial Parameters that apply to photographed subjects pertain but these works of
art also have their own additional challenges. Most notably, por­
Before the nineteenth century, most portraits were, almost by trait art does not provide what might be called a photographic
definition, depictions of people who were important in their own likeness but rather one that goes through a process of visual
worlds. But, as a walk through almost any major art museum will interpretation on the part of the artist.
show, a large number of these portraits from before the nine­ In the hope of addressing this challenge, the FACES re­
teenth century have lost the identities of their subjects through search team was formed in 2011 , seeking and receiving two
the fortunes of time. Traditionally, identification of many of these grants from the National Endowment for the Humanities, grants
portraits has been limited to often quite variable personal opin­ whose two years of support (2012-2014) allowed us to estab­
ion. FACES (Faces, Art, and Computerized Evaluation Systems) lish proof of concept of our project. In general, what this meant
proposes to establish the initial parameters of the application was identifying the issues, establishing the basic methodology,
of face recognition technology to the probable identification of applying the FACES algorithm that is the core of this methodol­
works of portrait art- this highly subjective aspect of art history ogy to a particular set of paradigms to establish the initial pa­
- while at the same time retaining the human eye as the final rameters of the technology (we use the term paradigm here to
2
arbiter. mean a logically chosen body of related images directed toward

267
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

a particular demonstrative end), and applying the FACES algo­ Practically speaking, this meant that they had to be from a his­
rithm to a body of chosen 'identifications'. torical period marked by its attention to naturalistic represen­
Identifying the issues was and is an ongoing process. tation and, in general, by artists known for such attention. We
The problems already mentioned that arise in traditional face found that portraits from Western Europe, fifteenth to early eight­
recognition applications (facial expression, age, facial hair, angle eenth century, suited our purposes, the images being selected
of pose, and so on) continue to be a challenge, though it does with as much control over variables as possible. The paradigms
seem that they will eventually be resolved. But, aside from this, themselves were conceived of as addressing a broad spectrum
it was immediately apparent that there was a lack of sufficient of issues, there being ultimately around seventy paradigms,
training data, that is, a lack of an established data set of images though many of these addressed identical issues for confirma­
(works of portrait art). Such a data set is used to establish stand­ tion purposes.
ards of probable match or non-match, the larger the data set, In regard to methodology more narrowly speaking, there
the greater the accuracy of the testing, such a lack being some­ are a number of analytical methods used in face recognition
thing that is not the case with most mainstream face recognition practices, all of which are not equally suitable in all contexts. Ini­
applications today. Another issue was the role of the style of the tial research indicated that, given the particular demands of this
individual artist in the creation of a portrait. But the key issue in first general study of the application of face recognition technol­
restoring lost identities to works of portrait art remained the initial ogy to portrait art, two key methods work best: the computation
one: how to test one portrait against another, both of which had of anthropometric distances and of local features. With anthro­
been subject to the subjectivity of artistic interpretation. pometric distances [Fig. 2], a network of very fine distances,
As to establishing the basic methodology, to begin with, giv­ both horizontal and vertical, are measured throughout the face.
en that our objective was to establish the initial parameters of the Through painstaking experimental trial and error, we established
application of face recognition technology to works of portrait a feature set (the most effective body of identifying features) of
art, it was important that the images that made up these para­ eleven possible measurements, such as the width between the
digms that were the basis of our research be visually suitable. eyes and the width of the mouth, in this way quantifying these
3
aspects of the essential structure of a given individual face.
With local features [Fig. 3], the characteristics of the edges of
features such as the corners of the eyes and the corners of the
Portrait Feature Space (PFS) - Match 4
mouth are assessed. Here, we identified a feature set of twen­
-Non-match
ty-two different features or landmark points, as they are called.
••• Standard match deviation
••• Standard non-match deviation These were marked on a generic face mesh that had been 'reg­
10 istered' onto the image being studied and the landmark points
9 were then quantified through the use of Gabor filters (a tech­
til
� 8 nique that analyzes edge gradients of these landmark points at
0
:g eight orientations along five scales, comprising a set with a total
:g 7
iii of forty different readings at each point).
:0 6
.c: Carefully selecting a body of images to form a data set, we
0
1ii 5 used this method ('feature extraction') to analyze several hun­
E
c dred portraits, among which some were known to be 'match'
0 4
e
"0
� images (known to share a common identity) and others known
ca 3
.c:
0
to be 'non-match' (known to not share a common identity).
1ii 2
:; The anthropometric distances and local features of each image
were computed and, when calculated for match images and

0 ����������� non-match, gave a number on a scale from 0 to 1 -the similarity


�������
0.4 0.5 0.6 0.7 0.8 0. 9 1.0 score - analogous to a percentage. The distribution of similarity

Similarity score
scores (match and non-match scores) that was produced from
this body of portraits is known as the 'portrait feature space'.
On the portrait feature space graph [Fig. 4], which gives the dis­
4. The FACES portrait feature space (the distribution of match tribution of similarity scores, the blue curve represents the distri­
and non-match scores). Additionally, the similarity score line is bution of match scores (images of known identities), and the red
marked with the result of the test between the unknown portrait curve indicates the distribution of non-match scores (images of
sometimes said to be of Monmouth [Fig. 1] and a known unknown identities). The area between the blue dotted lines de­
portrait of Monmouth [Fig. 5] notes the standard distribution of match scores; that is, the range

268
FACES: Faces, Art, and Computerized Evaluation Systems -A Feasibility Study of the Application of Face Recognition Technology...

5. Studio of Godfrey Kneller, «James Scott, Duke of Monmouth and Buccleuch», 1678, London, National Portrait Gallery,
NPG 5225. Photo: National Portrait Gallery

269
Conrad R udolph, Amit R oy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

where most of the match score images are found (there could be called p-value. This p-value is then subjected to well-known
match scores that are higher or lower). Conversely, the area be­ statistical tests to arrive at the decision. If the similarity score
tween the red dotted lines represents the standard distribution from the test lies within the standard match distribution (that
for non-match scores (again, there could be non-match scores is, between the blue dotted lines on the portrait feature space
that are higher or lower). Because the distribution of scores is graph, Fig. 4), the test indicates the probability - not the cer­
based upon a specific body of images, it will vary as the body tainty- of a match. Similarly, if the similarity score lies within the
of images varies. standard non-match distribution (between the red dotted lines),
With the portrait feature space established, let me contin­ it suggests the probability- not the certainty- of a non-match,
ue to describe our methodology as if two portraits were being granted that the images that were tested were relatively true like­
tested for similarity using the FACES algorithm: one, the test nesses and that there were no strongly mitigating factors (such
image, a portrait whose identity is uncertain [Fig. 1], and the as too great of an age difference, a face too strongly angled to
other, the reference image, a portrait known to be of a particu­ provide sufficient information, etc). If the similarity score is more
lar individual [Fig. 55]. The portrait whose identity is uncertain or less equally distant from the match and non-match distribu­
happened, in this case, to be by an unknown artist, and under­ tions (that is, between the innermost red and blue dotted lines
went feature extraction following the method described above. on the portrait feature space graph), the result is a 'no decision',
But the portrait whose identity is certain was by a known artist, meaning that the probability of a match or non-match is not sta­
Godfrey Kneller, and with a work by a known artist the element tistically conclusive.8
of style enters into in the analytical process. By style, we simply The process just described is the FACES algorithm.
follow Meyer Schapiro's understanding of style as 'the constant As to the results of the two portraits hypothetically being an­
form[... ] in the art of an individual or group'.6 alyzed with the FACES algorithm, when the portrait of uncertain
Anthropometric distances and local features are part of the identity thought by some to represent Monmouth [Fig. 1] was
body of characteristics that make an individual face individual, tested against a portrait known to represent Monmouth[Fig. 5],
and so they vary from face to face. At the same time, certain of we received a similarity score of 0.7204 (we also tested the por­
these features have an 'invariant' quality with individual artists. trait of uncertain identity against another portrait known to be of
That is, while some of them vary and so could contribute to the Monmouth, Fig. 69, receiving a similarity score of 0.69188). The
distinctive appearance of the sitters depicted by an individual similarity score of 0. 7204, represented by the dot on the simi­
artist, others tend to not vary and are part of the dynamic of the larity score line in Fig. 4, is one that is solidly within the match
style of the individual artist. In order to model- that is, to teach range (as is the score of 0.69188). And so, does the portrait of
the computer - the individual style of the artist, information the deceased man represent Monmouth?
(the anthropometric distances and local features) is extracted In the context of face recognition technology, this question
from a body of portraits of different sitters by the same artist. is more appropriately phrased as: how might the results of this
The information is then analyzed through a procedure known test be used in an art historical understanding of this portrait?
as random subspace ensemble classification. This procedure And the answer to that question is something that will be better
repeatedly selects random subsets of these extracted features understood after a fuller discussion of FACES.
from the body of portraits, ranking them in order of importance Another aspect in establishing proof of concept was the
from 'more invariant' to 'less invariant', potentially including fea­ application of the developing FACES algorithm to a particular
tures that might not be detectable to the human eye. In this, set of paradigms in order to establish the initial parameters of
random subspace ensemble classification strengthens the anal­ this technology to works of portrait art. While these paradigms
ysis of the information, giving a higher level of confidence in the were actually tested without any concern for order, they will
results of the analysis by allowing us to select different subsets be presented in this paper in the same logical sequence with
of the features in a probabilistic _way and see which one is the which FACES was originally conceived so as to guide the reader
most effective? through that same thought process.
With both images having undergone feature extraction - The essential unknown in this project is the relation of the
and with the image that was able to having undergone style element of artistic creativity or artistic interpretation to the work
modeling - the test image (the portrait whose identity is un­ of portrait art. Wishing to test the feasibility of the application of
known, Fig. 1) was now tested against the reference image (the face recognition technology to portraiture in the most straight­
portrait whose identity is known, Fig. 5), using a statistical hy­ forward way,_ we tested an analogue- an unmediated image of
pothesis test known as a Z-test. This determines whether the the subject, not a work of art - against the image of a three­
pair is a match or non-match (or if it is a 'no decision') by meas­ dimensional work of art that, in this case, physically approaches
uring the deviation of the similarity score from the mean value of the subject in form and size but that nevertheless partakes of the
the distribution under consideration in order to obtain a number subjectivity of artistic interpretation. And so we tested a cast of

270
FACES: Faces, Art, and Computerized Evaluation Systems -A Feasibility Study of the Application of Face Recqgnition Technology...

6. Portrait possibly after William Wissing, «James Scott, Duke of Monmouth and Buccleuch», c. 1683, London,
National Portrait Gallery, NPG 151. Photo: National Portrait Gallery

271
Conrad R udolph, Amit R oy-Chowdhury, Ramya Sr inivasan, and Jeanette Kohl

Urban VIII, against another sculptural portrait of the same indi­


vidual [Fig. 912, Fig. 1013], both around the same stage of the
individual's life and both depicted by the same artist, Gian Lor­
enzo Bernini. In other words, we proceeded with as much con­
trol over variables as possible, in the process obtaining a very
strong match score of 0. 7572. Although the similarity between
these two portraits might seem 'self-evident', to quote one 'sup­
portive critic' of FACES, the point is not that such a test can
tell us what we already know but that the technology can work,
that it can provide an objective, quantified basis for visual rela­
tionships between two works of art - and this is something that
is best done with images whose resemblance is not already in
question.
Incrementally, we broadened our tests- too involved to fully
detail here - introducing a similarly controlled but wide-ranging
series of systematically chosen variations extending from more
controlled paradigms to less controlled (that is, more challeng­
ing) ones. These included the same stage of an individual's life
but by different artists (for example, Cardinal Scipione Borghese
by both Bernini and Giuliano Finelli; strong match score of
0.7324)14, different stages of an individual's life by the same art­
ist (Urban VIII, although only ten years apart, by Bernini; very
solid match score of 0.7003)15, and different stages of an indi­
vidual's life by different artists (Pope Alexander VII, once more
only ten years apart, by Bernini and Melchiorre Cafa; a solid
match score of 0.6989)16 - all in three-dimensional imagery.
Again, these tests do more than confirm what might seem to
be self-evident. They demonstrate that this technology has the
ability - despite the basic similarity of most human faces - to
recognize the same subject even when rendered by different art­
7. Possibly by Orsino Benintendi, «Lorenzo de' Medici», death ists, each no doubt very conscious of his own personal style. Put
mask cast (originally gilded), 1492, Florence, Accademia another way, they demonstrate that a particular body of measur­
Toscana di Scienze e Lettere 'La Colombaria', on loan to the able features may at times be retained by different artists in their
Museo degli Argenti. Photo: Museo degli Argenti depictions of the same sitter. And they show that the technology
can accept at least moderate age difference - not only in works
by the same artist but, which is more challenging, by different
the death mask of a known individual against an identified sculp­ artists.17 Together, these tests tentatively attest to the general ac­
0
tural portrait of the same individual: Lorenzo de' Medici [Fig. 71 , ceptance by this technology of the normal variables that occur
Fig. 811]. In this, we received a match score of 0.7114, a rather in the genre of portraiture and suggest a strong foundation for
strong score and one that indicates that a particular work of further development.
portrait art could in fact carry what might be called an objective Expanding our examination, we tested two-dimensional
set of facial signifiers (the artistic rendering of anthropometric imagery (painting and drawing), first simply considering two
distances and local features) while at the same time being sub­ two-dimensional images of the same subject by the same artist
ject to the subjectivity of artistic interpretation - all of this being (self-portraits by Bernini; 0.7342, a strong match score)18, and
dependent, of course, upon the method, ability, and intentions then mixing media by testing a number of sculpture vs. painting
of the individual artist. (that is, three-dimensional vs. two-dimensional) paradigms (for
Leaving the relative security of the analogue and work of art example, a sculpted bust and a painting of Urban VIII, both by
pairing, we tested a paradigm of exclusively three-dimensional Bernini; 0.7104, a solid match score).19 In this, we employed
works of art; that is, we tested two works both of which were now a systematic series of distinctions similar to those already men­
subject to the subjectivity of artistic interpretation. More specifi­ tioned (same artists, different artists, same ages, different ages,
cally, we tested a sculptural portrait of a known individual, Pope and so on). The testing of two-dimensional imagery against

272
FACES: Faces, Art, and Computerized Evaluation Systems -A Feasibility Study of the Application of Face Recognition Technology. . .

8. Probably after a model by


Andrea del Verrocchio and
Orsino Benintendi, «Lorenzo
de' Medici», probably 1478-1521,
Washington, National Gallery
of Art. Photo: National Gallery
of Art

two-dimensional imagery is straightforward enough: images human face or a photographed work of sculpture- painting and
of the same medium tested against each other. But the testing drawing undertake the same process of transformation of the
of three-dimensional against two-dimensional imagery raised three-dimensional into the two through the thought process of
other questions, at least initially. For while photography ren­ the artist. Something that might potentially, at least, create chal­
ders a three-dimensional subject into a two-dimensional image lenges for facial recognition technology- and something that we
automatically, through technology - whether a photographed had to be aware of- our tests found that this was not an issue.

273
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

9. Gian Lorenzo Bernini, «Pope Urban VIII», c. 1632, Rome, Galleria Nazionale d'Arte Antica, Palazzo Barberini. Photo: Art Resource

274
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

10. Gian Lorenzo Bernini, «Pope Urban VIII», c. 1632, Ottawa, National Gallery of Canada. Photo: National Gallery of Canada

275
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

11. Francesco Laurana, «Battista Sforza», c. 1474, Florence, Museo Nazionale del Bargello. Photo: Art Resource

276
FACES: Faces, Art, and Computerized Evaluation Systems -A Feasibility Study of the Application of Face Recognition Technology...

Identifications with FACES

The final aspect that concerned us in establish­


ing proof of concept was the application of the
developing FACES algorithm to a series of inter­
esting and sometimes important 'identifications'.
In this, we do not claim that we are providing the
incontestable identity of the sitter in question,
but only that the similarity score of a given pair of
portraits does or does not fall within the distribu­
tion of match or non-match scores of the portrait
feature space. Let us run through a number of
our identification paradigms, sometimes merely
as examples of the potential of FACES, some­
times drawing attention to various strengths or
weaknesses.
For some time, art historians have felt that
the posthumous bust of Battista Sforza by Fran­
cesco Laurana in the Bargello (c. 1474; Fig. 1120)
was modeled after a death mask cast also by
Laurana now in the Louvre (1472; Fig. 1221). The
similarity score of 0.7116 from our test indicates
the strong probability of a match, regardless of
the obvious challenges to such a comparison.
This suggests that, as was thought, the mask
was taken from Battista, and supports as well
the idea that the cast was quite closely followed
by Laurana as a model - rather than, say, fol-
lowing Piero della Francesca's profile portrait of 12. Francesco Laurana, «Battista Sforza» (?), death mask cast, c. 1472, Paris,
Battista, though it or other images most certainly Musee du Louvre, RF 1171. Photo: Art Resource
may have been referred to by the artist in the
course of his work.22 A score strongly within the
match range of 0.727 was also returned for Botticelli's Portrait of around 1590, possibly shortly after Galileo had been offered
a Lady at the Window (c. 1475; widely thought to be a render­ a chair in mathematics at the University of Pisa [Fig. 1526]. When
ing of Smeralda Brandini; Fig. 1323) and Verrocchio's Lady with tested against a chronological spectrum of eight other known
Flowers (c. 1475; Fig. 1424), the two portraits sometimes being portraits of Galileo, the results neatly broke down into consist­
suggested by some to represent the same sitter, thus lending ently decreasing similarity scores: first, within the match range
objective support to this position despite the two distinctly dif­ for the chronologically three closest likenesses (1601-c. 1612);
ferent personas conveyed in the images. Since neither portrait then, within the 'no decision' range for the next two (1624); and,
has a certain identity, the probable match is relative rather than finally, within the non-match range for the final three (1635-
absolute, but a probable match nonetheless. Perhaps of a little 1640).27 While age remains a challenge for FACES and requires
less prominence but of no less interest is the portrait by Michiel more research (this is also the case with photographed human
van Musscher of his family, where a similarity score of 0.708 faces), age differences of around ten years or so have not been
was given between a person who is obviously the artist's wife - too much of an obstacle. If the test image here does in fact date
thought by some to depict the artist's first wife, Eva Visscher, and from around 1590 and does depict Galileo, this body of portraits
by others his second wife - and a known portrait of Eva by him, would seem to represent something of an exception in the age
an identification that can affect the immediate understanding of question in that strong similarity scores were returned for por­
the painting.25 traits with an approximately twenty-four year age difference.28
In a somewhat different vein is the portrait, sent to us by the (A particular body of portraits of Isaac Newton shows an even
Italian astronomer Paolo Molaro, of what he believes may be broader spectrum, although it is not so consistent in its scores
the earliest known likeness of Galileo Galilei, painted perhaps as the Galileo paradigm).29

277
C onrad R ud olph, Amit R oy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

13. Botticelli, «Portrait of a Lady


at the Window (Smeralda
Brandini?)», c. 1470-1475,
London, Victoria and Albert
Museum. Photo: Victoria and
Albert Museum

278
FACES: Faces, Art, and Computerized Evaluation Systems -A Feasibility Study of the Application of Face Recognition Technology...

14. Verrocchio, «Lady with Flowers», c. 1470-1475, Florence, Museo Nazionale del Bargello. Photo: Art Resource

279
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

15. Artist unknown, «Galileo Galilei» {?),c. 1590, Trieste, private collection

280
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

30
represent Robert Devereux, second earl of Essex. Essex's like­
ness has survived in a fairly reasonable number of paintings and
miniatures. One challenge in identifying the sitter of Young Man
Among Roses as Essex is that Essex grew a beard around 1596,
not only obscuring his facial features in portraiture with regard to
traditional identification but also making it more difficult for the
effective use of face recognition technology. Among the body of
portraits of Essex, curator Elena Greer feels that the most useful
one for comparison with Young Man Among Roses is the depic­
tion of Essex by William Segar of around 1590, which is around
the same date as Young Man Among Roses and before Essex
31
grew his beard. Indeed, when Young Man Among Roses was
tested against the Segar portrait, the score that was returned
was a rather strong 0.72. Furthermore, when tested against
the Hilliard miniature at the Metropolitan Museum in New York
(1588), believed to probably represent Essex, the result was
32
a score of 0.7. And when tested against a Hilliard miniature
at the National Galleries of Scotland in Edinburgh, thought by
some to depict Essex, the return was a score of 0.705, with this
Edinburgh miniature also receiving a score against the Segar
33
portrait of 0.71. Scores in the non-match range were returned
for all tests of Young Man Among Roses against the bearded por­
trayals of Essex, as well as for one miniature of a clean-shaven
figure generally thought to be Essex of c. 1587 attributed to the
34
studio of Hilliard. Generally speaking, while a non-match can
be the result of a number of factors aside from that of different
identities -such as issues of facial hair, poor likeness, age differ­
ence, angle view, and so on, just as with photographed human
faces - a match is unlikely to be coincidental. And when there
is a pattern of match and non-match images as we received
for this paradigm - where the beard of the sitter could easily
account for all of the non-match portraits but one - the weight
of the matches tends to suggest that the miniature thought by
some to represent a beardless or largely beardless Essex from
the studio of Hilliard is probably best understood as simply
a less true (or less conforming) likeness than the other beard­
less or largely beardless depictions, an exception that does not
tend to discredit the match scores. These tests, therefore, pro­
vide a body of quantitative evidence derived from the images
themselves that strongly supports the traditional supposition
that Young Man Among Roses represents Essex.
16. Nicholas Hilliard, «Young Man Among Roses», c. 1587, London, The results of FACES are only as dependable as the images
Victoria and Albert Museum, P.163-191 0. Photo: Victoria and tested, and some paradigms gave seemingly undependable re­
Albert Museum sults. In a paradigm in which we tested the only known 'portrait'
35
of Anne Boleyn - the 'Moost Happi' medal - the results were
such as to suggest that the test image itself did not provide suf­
One paradigm that traverses some of the potentially ficient portrait data for a reliable test.
complex terrain that one might encounter in the use of this A variation on this same situation at least seems to be the
technology is that of Nicholas Hilliard's Young Man Among Ros­ case with the tangled web of portraits that different proponents
es (c. 1588; Fig. 16), said by Roy Strong to be 'perhaps the have claimed at one time or another portray William Shakespeare
36
most famous miniature ever painted', and generally thought to (1564-1616). Perhaps the leading authority on the portraits of

281
Conrad R udolph, Amit Roy-Chowdhury, Ramya Sr inivasan, and Jeanette Kohl

17. Cast of the sculpture, probably by


Garrat Johnson (Gheerart Janssen)
the Younger, of «William Shakespeare»,
c. 1620, Stratford, Holy Trinity Church,
cast NPG 1735. Photo: National Portrait Gallery

18. Martin Droeshout, «William Shakespeare»


(Droeshout engraving, First Folio,
first state, published 1623).
Photo: Folger Shakespeare Library

Shakespeare, Tarnya Cooper, believes that only two extant por­ body of anthropomorphic distances and local features consist­
traits provide us with a reasonable sense of Shakespeare's like­ ent enough to allow the images to return similarity scores that
ness, the Stratford bust (c. 1620, probably by Garrat Johnson are able to support a clear match or non-match- and, beyond
37
the Younger; Fig. 17 ) and the Droeshout engraving (published that, to return such scores from any of the rest of the main body
38
1623, by Martin Droeshout; Fig. 1 8 ) Neither of these depic­
. of claimed Shakespeare portraits, a process we have seen is
tions, however, is an accomplished work of portrait art, some­ possible in principle from earlier, 'self-evident' paradigms.
thing that does not contribute to reliable test results with FACES. It is Cooper's opinion that the Stratford bust and Droeshout
At the same time, while the Droeshout engraving was said by engraving- both posthumous to Shakespeare - were based on
40
Ben Johnson to represent the face of Shakespeare well, the a now lost previous portrait for which Shakespeare sat. It is
context of this statement- a partially cliche-based verse dedica­ also her opinion that the so-called Chandos portrait is the most
tion as opposed to some less contrived discourse - suggests likely of all the claimed Shakespeare likenesses to have been
that this 'opinion' need not necessarily be taken as an objective made in the author's lifetime (c. 1600-1610, attributed to John
41
avowal that the engraving is a true likeness as understood in Taylor; Fig. 19 ). But when we look at the FACES test returns
39
this study. T he concern here is not whether either of these por­ for this body of portraits, we find a welter of ambiguous 'no de­
traits generally looked like Shakespeare to his contemporaries cision' associations between the many portraits. T he problem
but whether their creators were capable of rendering a sufficient arises in part from both the features and the quality of many

282
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

19. Attributed to John Taylor, «William Shakespeare»(?) (Chandos portrait}, c. 1600-1610, London, National Portrait Gallery, NPG 1.
Photo: National Portrait Gallery

283
Conrad R udolph, Amit R oy-Chowdhury, Ramya Sr inivasan, and Jeanette Kohl

20. Artist unknown, identity of sitter in question, c. 1570, London, National Portrait Gallery, NPG 96. Photo: National Portrait Gallery

284
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

of the portraits: beards and mustaches that limit information, - the results were almost startling, showing four scores in the
copies or possible copies - even copies of copies - that may match range (three Clouets and the Hilliard) and four in the 'no
contribute to a degradation of information, different levels of ac­ decision' range, some of the latter being quite close to the match
complishment on the part of the various artists that potentially range (including three Clouets)- all this despite the differences
limit or misrepresent information, and overpainting and cleaning in style in the representation of the face.50 (The score in the non­
that also potentially limit or misrepresent information. But there match range was with her tomb sculpture in Westminster Abbey.)
is more to the Gordian knot of Shakespeare portraits than this. The second portrait was that thought by some to be of
There is also apparently another, fundamental difficulty inherent Monmouth with which this paper opened and which the exhibi­
in this body of images. For, from the very beginning, as it were, tion described as of an unknown man [Fig. 1]. As mentioned,
we see that while the Stratford bust is loosely associated through when tested against two portraits known to represent Monmouth
scores in the 'no decision' range - that is, scores that could [Figs 5, 6], it received scores in the match range of 0.7204 and
be said to be not non-matches - with the Droeshout engrav­ 0.6918 (it also received a score in the 'no decision' range of
ing (0.663) and with the (heavily overpainted) Chandos portrait 0.6772 when tested against a relatively unaccomplished mezzo­
(0.681), it is also similarly loosely associated with the Janssen tint copy of an earlier portrait51). While these are strong scores,
portrait (0.643).
42 The Droeshout engraving has the same loose the eyes of the deceased man are closed and the face is at an
'no decision' associations, but it is additionally loosely associ­ angle, both of which reduce the amount of information currently
ated with the so-called Cobbe portrait of Shakespeare (0.654, computable by the FACES algorithm.52
'no decision' range).43 At the same time, the Chandos portrait And so, does this portrait represent Monmouth -do any of
follows this same pattern of initial, loose associations but now the portraits that received scores supporting a match represent
with an increasing number of additional loose, 'no decision' the individuals against whose known portraits they were tested?
associations.44 Of all the portraits in this paradigm (in which
scores in the 'no decision' range far outnumbered those in the
non-match range), the only two to receive a score in the match FACES: Concluding Remarks
range (0.692} were the Janssen portrait and a portrait of Thom­
as Overbury (1581-1613) by Marcus Gheeraerts the Younger A great portrait is more than the sum of its parts and certainly
(c. 1610).45 If it is true, as is generally believed, that the Strat­ cannot be reduced to a discrete series of objective factors that
ford bust and Droeshout engraving offer reasonable likenesses are readily subject to the 'superior' discrimination of the com­
of the historical figure of Shakespeare, one unavoidable if not puter. However, setting aside what in this case might rightly or
very satisfying conclusion seems to be - aside from the other wrongly be called 'extraneous' elements- elements sometimes
problems of beards, condition, quality, and so on-that the intro­ associated with the portraiture of a particular artist, including the
duction of the very similar physiognomy of at least one different general composition, setting, pose, facial expression, lighting,
individual (possibly Thomas Overbury via the Janssen portrait, skin treatment, hair, drapery, finish, and so on - setting aside
as has been previously suggested46) into the body of claimed these elements, the face remains. Certainly more elusive than
Shakespeare portraits is compounding the lack of clarity of this the other elements, the face can nevertheless still be quantified
already challenging body of images.47 Even so, the large num­ in a way not so dissimilar to the traditional canons employed
ber of scores in the 'no decision' range - which indicate that by artists in the depiction of the human being from the ancient
there are some similarities but not enough to meet generally Egyptians to the Classical Greeks to the artists of the Renais­
recognized standards - suggests that, as this technology ad­ sance. With this understanding, at the end of two years of re­
vances, future research on this and perhaps a broader body of search, we have established proof of concept of the application
portraits may very well contribute to a more secure sense of the of face recognition technology to works of portrait art. This in­
likeness of the greatest writer in the English language. cludes identifying the issues, establishing the basic methodolo­
Finally, the relatively recent exhibition at the National Portrait gy, applying the FACES algorithm (the core of this methodology)
Gallery in London, Imagined Lives: Portraits of Unknown People, to a carefully chosen set of paradigms in order to determine the
in which a number of portraits whose identities have been con­ initial parameters of the technology, and applying the FACES
tested over the years were revisited, suggested to us that we algorithm to a body of selected 'identifications'. In these first
look at two of the more interesting images.48 steps, there is no more reason to expect completion or perfec­
The first was the portrait of a woman thought at one time to tion than there is with any other new digitally based technology,
represent Mary Queen of Scots, at another to depict her mother, all of which typically go through long periods of development.
Mary of Lorain, and currently simply to be unidentified (c. 1570; As with seemingly everything digital, this is a work in progress,
Fig. 2049). When tested against eight other portraits known to be a procedure that will certainly be refined in probably every area,
of Mary - including six by Frangois Clouet and one by Hilliard including the most challenging.

285
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

The FACES algorithm does not operate like the make-be­ understanding of the artistic process in cases where casts of
lieve computer processes a person sometimes sees in popular one kind or another were used. But the FACES algorithm could
media. Nor is it for every portrait. As with photographed human also be applied to a study of adherence or non-adherence to
faces, works of portrait art are subject to variations in facial ex­ widely recognized artistic canons, formal or informal, as well as
pression, age, facial hair, angle of pose, lighting, and the very to the identification of variations in the practice of an individual
identity of the images tested. artist: over time, with different subjects, with different genres, af­
However, unlike photographed human faces, works of por­ ter exposure to external influences, and so on. It certainly might
trait art are also subject to the subjectivity of artistic interpreta­ eventually be possible to more confidently determine probable
tion, whether simply from issues of style or from something more bodies of work of anonymous artists - including in the so-called
pronounced, such as a consciously more flattering rendering of minor arts- or to identify probable unrecognized works of known
the sitter (queens, in particular, appear to be a productive area artists. And it might be possible to study difference in larger bod­
of future research in this regard53). But there are also other fac­ ies of works- art historical 'big data'?- for instance, a particular
tors involved, including artistic quality, with renderings of less body of art from Italy during a particular time frame examined
mimetic accomplishment affecting the readings of portraits no in relation to a corresponding particular body from Northern
less in original images than in copies, which sometimes, though Europe during a corresponding time frame. The FACES team
not always, are susceptible to distinct degradation of detail.54 wanted, but was unable for reasons of time, to move out of the
Post-production interventions (restoration, overpainting, exces­ Western sphere: for example, to take up an unusually large body
sive cleaning, repairs, etc.) are a constant and serious concern. of portraits that appear to depict the same Moche noble through­
And it may even be that, on occasion, the use of different media out the course of his life (Andes, c. 100-800 AD).56 The basic
- terracotta as opposed to marble, chalk in contrast to oil, for ex­ concept of FACES could even be applied in an altered form to
ample- which human perception reads as more or less identical ornament of all kinds and especially to certain minor variations in
are read more precisely by these tests (apparently because of medieval (at least) architectural detail - by which I do not mean,
the sometimes greater exactitude of particular media in detailing for example, different types of moldings but rather more or less
local features), this also being an area that might benefit from the same molding rendered in slightly different ways according
future research.55 Finally, the human element in the processing to the particular practice of the individual mason - something
of these tests may also sometimes result in very minor variations that potentially might reveal a wealth of information about build­
in scores due to extremely slight differences in the registering of ing processes, building history, and architectural identities in
points of measurement on the facial meshes. buildings that are sometimes national treasures. Perhaps the ini­
In all this, we recognize that the style of the individual artist tial impetus to FACES might be one of its greatest challenges:
is not a simple issue, being complicated by any number of fac­ working out as much as possible the identities of particular bod­
tors from the work of contemporary assistants to interventions of ies of miniature portraits of uncertain identity from the sixteenth
later restorers, and future work in this area will undoubtedly take century on, portraits that were often given as part of social and
the subject further. But, at this point, the test results suggest that political networks and that might help recover lost networks at
local features carry a distinctly greater invariant quality than do different levels of society, from the local to the national to the
anthropomorphic distances. It would seem that these particular international. These people, many of them forgotten now, were
patterns of invariant features for individual artists are part of the important in their own time. By restoring their identities we would
dynamic of the style of an individual artist. In our work on this, be doing more than just contributing to these portraits as works
which amounts to only the first steps in approaching a very com­ of art, we would be restoring historical 'documents' that can tell
plex subject, we emphatically do not claim to identify the style us more about our own past - just as much, if handled right,
of individual artists. But we are able to identify certain patterns as any historical document, the value of which is unquestioned.
that appear in the particular bodies of portraits that we tested, The application of the FACES algorithm should be made
patterns that could be of great interest to scholars who would a part of standard curatorial practice where suitable, whether
like to know more about the different ways- now tending toward during initial acquisition or later study. Results based on agreed
'invariants', as it were, now 'variants' - in which these artists ap­ upon standards should be made available online: while an anal­
proached their various subjects [cf. Fig. 21]. ogy between face recognition and fingerprints is imperfect,
The use of the FACES algorithm should not be thought of as it should be pointed out that, with such technologies, their effec­
limited to facial recognition in the sense of identification alone. tiveness increases exponentially with the size and organization
We have already mentioned using FACES to study difference of the data set. The use of technology in the study of works of art
in the depiction of various bodies of sitters, queens being one - X-rays, infrared reflectography, radio spectroscopy, chemical
among a number of potential subjects. And we have suggest­ analysis, and so on - is not going away. Indeed, it is becoming
ed that FACES might sometimes be of help in getting a better an ever more firmly established part of the curatorial endeavor.

286
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

Characteristic anthropometric distances Characteristic local features


in decreasing order of importance: in decreasing order of importance:

Algardi: 4,8,3,7,2,11 Algardi: 8,5,18,13,6,1,2,3,14,11,19,4,17,16


Bernini: 1,10,7,5,8,6,1 Bernini: 2,3,11,8,12,6,4,20,10,15,14,1,18,17
Clouet: 5,11,2,7 Clouet: 6,3,2,19,22,14,11,9,5,8
Holbein: 2,8,11,3 Holbein: 11,8,5,7,17,22,16,19,4,1
Kneller: 2,1,8,9,10,5,4,3,6 Kneller: 5,20,2,13,4,1,9,17,10,16,18,11,19,6
Van Mierevelt: 4,6,11,7,3,1,8,9 Van Mierevelt: 2,5,8,14,21,17,9,11,13,21
Van Musscher: 9,7,3,11,2 Van Musscher: 8,15,5,3,4,1,9,11,21

List of anthropometric distances List of local features

1: distance between forehead tips 1: forehead tip (left)


2: distance between forehead center and chin bottom 2: forehead tip (right)
3: distance between nose top and bottom 3: forehead center
4: distance between points on temples 4: chin bottom
5: distance between chin ear corners 5: nose top
6: distance between points on chin 6: nose bottom
7: distance between pupils 7,8: points on temple (left, right)
8: distance between cheekbones 9,10: chin ear corners (left, right)
9: distance between mouth corners 11,12: points on chin (left, right)
10: width of nose 13,14: cheekbones (left, right)
11: distance between forehead center and nose bottom 15,16: mouth corners (left, right)
17,18: pupils (left, right)
19,20: left eye corners (left, right)
21,22: right eye corners (left, right)

21. List of FACES artists' characteristics in regard to anthropometric distances and local features

None of this technology has taken the place of the eye, knowl­ What could be more complex than the human face -the in­
edge, or experience of the conservation scientist, curator, or art dex of the mind, the window of the soul, and, at times, an impen­
historian. Rather, it has simply drawn attention to previously un­ etrable mask? Computers can extract features in a quantitative
noticeable or unconfirmable information by contributing catego­ way that, generally speaking, are usually only intuitively sensed
ries of objective data to be used in conjunction with larger bodies with human vision, thus providing new categories of information
of information to which the researcher already has access. that were not truly extant before for the art historian or curator to
And so, does the portrait of the deceased man represent take into consideration in their research. T hey can, sometimes,
Monmouth? While recognizing that the angle of the face and provide an objectifying factor in otherwise often very subjective
the closed eyes limit measurable information, the similarity score cases. While we are insistent that this technology does not prove
returned for this portrait does lend new and undeniable support the identity of its subjects - either absolutely or relatively - its
to previous opinion that the figure depicts the young, self-pro­ results, when probable, match what is known with a given un­
claimed anti-Catholic rebel, in the process contributing to an un­ known. Typically, this is something that is unlikely to be acciden­
derstanding of this portrait as a unique image that would seem tal, and so the results tend to be believable. It is in this sense
to have served an unusual purpose at a dramatic and transitional that this technology, when effectively used, can not only help
period in British history, the whole story told- to contemporaries, solve old and vexing art historical problems but also - through
at least-through Monmouth's pose, perhaps the small piece of the possible restoration of lost identities-help deepen our view
cloth at his throat, and certainly the expression on his face. of the history of a given time and place.

287
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

FACES was conceived of at the University of California, Riverside, in the between ear corners at chin, between points on chin, between irises, be­
spring of 2011, the team being composed of Conrad Rudolph (project tween cheekbones, between mouth corners, between forehead center
director). Amit Roy-Chowdhury (expert on face recognition technology), and nose bottom, and the width of nose.
Ramya Srinivasan (Graduate Student Researcher), and Jeanette Kohl 4 Left forehead tip, right forehead tip, forehead center, chin bottom, nose
(an authority on portraiture and the face in art). The FACES project is
top, nose bottom, left point on temple, right point on temple, left chin
a collaboration of the humanities (art history) and the sciences (com­
ear corner, right chin ear corner, left point on chin, right point on chin,
puter science). This article presents FACES from the point of view of
left cheekbone, right cheekbone, left mouth corner, right mouth corner,
the humanities, that is, how this technology generally works, what the
left iris, right iris, left corner of left eye, right corner of left eye, left corner
parameters of its application to portrait art are at this time, what its ad­
of right eye, and right corner of right eye.
vantages are, and so on. For the computer science basis of the study,
see R. Srinivasan, C. Rudolph, and A. Roy-Chowdhury, 'Computerized 5 James Scott,Duke of Monmouth and Buccleuch, studio of Godfrey Knel­
Face Recognition in Renaissance Portrait Art', Signal Processing Maga­ ler, 1678; London, National Portrait Gallery, NPG 5225.
zine, 32:4, July 2015, pp. 85--94. These two papers are meant to operate
6 M. Schapiro, 'Style', in Aesthetics Today, ed. by M. Philipson, Cleveland,
as a pair.
1961, p. 81. For a study of the quantification of style in the drawings of
For their much appreciated advice, we would like to thank Lucy
Pieter Bruegel the Elder, see J. M. Hughes, D. J. Graham, and D. N. Rock­
Churchill, Tarnya Cooper, Anne-Lise Desmas, Christopher Donnan,
more, 'Quantification of Artistic Style through Sparse Coding Analysis in
Lee Durkee, Robert Gerhardt, Bryce Kretschman, Alison Luchs, Paolo
the Drawings of Pieter Bruegel the Elder', PNAS [Proceedings of the
Molaro, Beth Morrison, Stella Nair, Steven Ostrow, and especially the
National Academy of Sciences of the United States of America). 2010,
National Endowment for the Humanities for the two grants that made
107(4). DOI:10.1073/pnas.0910530107.
this project possible.
7 For a more detailed discussion of this, seeR. Srinivasan, C.Rudolph, and

A. Roy-Chowdhury, 'Computerized Face Recognition in Renaissance


London, National Portrait Gallery, NPG 1566.Death, Passion and Poli­ Portrait Art', Signal Processing Magazine, 32:4, July 2015, pp. 85-94.
tics: VanDyck's Portraits of Venetia Stanley and GeorgeDigby, ed. by For an attempt to train computers to recognize schools of art, see Shamir
A. Sumner, London, 1995, cat. 48; National Portrait Gallery, Imagined and Tarakhovsky, 'Computer Analysis of Art'.
Lives: Portraits of Unknown People, London, 2011, p. 92.
8 For a more detailed discussion of this, see Srinivasan, Rudolph, and
2 To the best of our knowledge, FACES is the first systematic study of
Roy-Chowdhury, 'Computerized Face Recognition'.
the application of face recognition technology to works of portrait art,
although there has been a wide array of studies of art employing com­ 9 James Scott, Duke of Monmouth and Buccleuch, possibly after William

puter analysis. For a review of the literature in general, see D. G. Stork, Wissing, c. 1683; London, National Portrait Gallery, NPG 151.

'Computer Vision and Computer Graphics Analysis of Paintings and 10 Lorenzo de' Medici, stucco death mask cast (originally gilded), possibly
Drawings: An Introduction to the Literature', LNCS [Lecture Notes in by Orsino Benintendi, 1492; Florence, Accademia Toscana di Scienze e
Computer Science). 5702, 2009, pp. 9-24. For a study using face rec­
Lettere 'La Colombaria', on loan to the Museo degli Argenti. K. Lange­
ognition technology in particular, C. W. Tyler, W. Smith, and D. G. Stork, 1h 1h Centuries, 3 vols, Florence,
dijk, The Portraits of the Medici: 15 -18
'In search of Leonardo: Computer-based Facial Image Analysis of
1981-1987, cat. 74:25; Christiansen and Weppelmann, The Renaissance
Renaissance Artworks for Identifying Leonardo as Subject', in Human
Portrait, cat. 56.
Vision and Electronic Imaging, ed. by B. E. Rogowitz et a/. 17, SPIE Pro­
11 Lorenzo de' Medici, bust probably after a model by Andrea del Ver­
ceedings, 8291, 2012. For an attempt at the application of this technol­
ogy to style, see L. Shamir and J. A. Tarakhovsky, 'Computer Analysis of rocchio and Orsino Benintendi, probably 1478-1521; Washington, Na­
Art', Journal on Computing and Cultural Heritage, 5, 2012, article 7. For tional Gallery of Art. Langedijk, The Portraits of the Medici, cat. 74:28;
a discussion of pre-digital face recognition studies, H. Hammerschmidt­ A. Luchs, 'Lorenzo from Life?: Renaissance Portrait Busts of Lorenzo
Hummel, The True Face of William Shakespeare: The Poet'sDeath Mask de' Medici', The Sculpture Journal, 4, 2000, pp. 6-23. As pointed out by
and Likenesses from Three Periods of His Life, London, 2006, pp. 47-65. Luchs, /estoration should be taken into account in any future analysis,
For some recent work on portraiture that has been useful in this pa­ something beyond the scope of this initial testing; on the restoration,
per, M. Painton, Hanging the Head: Portraiture and Social Formation in see G. Bitossi, M. Mauro, and D. Rossi, 'La maschera funebre di Lorenzo
Eighteenth-century England, New Haven, 1993; R. Brilliant, Portraiture, il Magnifico: Caratterizzazione chimico-fisica della stucco originale e dei
London, 1990; The Image of the Individual: Portraits in the Renaissance, successivi restauri', in Lo stucco: Cultura, tecnologia, conoscenza, ed.
ed. by N. Mann and Luke Syson, London, 1998; E. Pommier, Theories by G. Biscontin and G. Driussi, Venice, 2001, pp. 519-526.
du portrait:De Ia Renaissance aux Lumieres, Paris, 1998; Portrat, ed. by 12 Pope Urban VIII, by Bernini, c. 1632; Rome, Galleria Nazionale d'Arte An­
R. Preimesberger et at., Berlin, 1999; P. Zitzlsperger, Gianlorenzo Bernini.
tica, Palazzo Barberini. Zitzlsperger, Gianlorenzo Bernini.Die Papst- und
Die Papst- und Herrscherportrats. Zum Verhiiltnis von Bildnis und Macht,
Herrscherportrats, pp. 80-87, cat. 12; Bernini and the Birth of Baroque
Munich, 2002; M. Porter, Windows of the Soul: The Art of Physiognomy in
Portrait Sculpture, ed. by A. Bacchi, C. Hess, and J. Montagu, with the
European Culture 1470-1780, Oxford, 2005; P. Rubin, Images and Iden­
assistance of A.-L. Desmas, Los Angeles, 2008, p. 288, list no. A18a.
tity in Fifteenth-Century Florence, New Haven, 2007; K. Christiansen and
S. Weppelmann, The Renaissance Portrait: From Donatello to Bellini, 13 Pope Urban VIII, by Bernini, c. 1632; Ottawa, National Gallery of Cana­
New York, 2011 ; Similitude. Ahnlichkeit in der Kunst des Mittelalters und da. Zitzlsperger, Gianlorenzo Bernini.Die Papst- und Herrscherportrats,
der Friihen Neuzeit, ed. by J. Kohl et a/., Munich, 2012; H. Belting, Faces: pp. 80-87, cat. 8; Bernini and the Birth of Baroque Portrait Sculpture,
Eine Geschichte des Gesichts, Munich, 2013. cat. 2.5.

3 T he distances between forehead tips, between forehead center and 14 Cardinal Scipione Borghese, by Bernini, 1632; Rome, Galleria Borghese,
chin bottom, between nose top and bottom, between points on temples, CCLXVI; Bernini and the Birth of Baroque Portrait Sculpture, cat. 4.1.

288
FACES: Faces, Art, and Computerized Evaluation Systems -A Feasibility Study of the Application of Face Recognition Technology...

Cardinal Scipione Borghese, by Giuliano Finelli, 1631-1632; New York, 23 Portrait of a Lady at the Window, Smeralda Brandini (?), by Botticelli,
Metropolitan Museum of Art; Bernini and the Birth of Baroque Portrait c. 1470-1475; London, Victoria and Albert Museum. Christiansen and
Sculpture, cat. 5.4. Weppelmann, The Renaissance Portrait, cat. 14.

15 The Galleria Nazionale d'Arte Antica Urban VIII, see note 12. Pope Urban 24 Lady with Flowers, by Verrocchio, c. 1470-1475; Florence, Museo
VIII, by Bernini, 1640-1644; Spoleto, Cathedral; Bernini and the Birth of Nazionale del Bargello. A. Butterfield, The Sculptures of Andrea del
Baroque Portrait Sculpture, p. 289, list no. A19. Verrocchio, New Haven, 1997, cat. 15. The Lady with Flowers is also
sometimes thought to represent Ginevra de' Benci (by Leonardo da
16 Pope Alexander VII, by Bernini, 1657; private collection; Bernini and the
Vinci, c. 1474-1478; Washington National Gallery), a test that resulted
Birth of Baroque Portrait Sculpture, cat. 6.6. Pope Alexander VII, by Mel­
in a 'no decision' score of 0.649. On this portrait, see Butterfield The
chiorre Cafa, 1667; Ariccia, Palazzo Chigi; Bernini and the Birth of Ba­
Sculptures of Andrea del Verrocchio, pp. 90-94.
roque Portrait Sculpture, pp. 263-265.
25 Family of the Artist, by Michiel van Musscher, 1694-1701; Antwerp, Royal
17 For more on aging, see note 29.
Museum of Fine Arts; lnv. 739. Eva Visscher, by Van Musscher, c. 1685;
18 Bernini, self-portrait, c. 1623; Rome, Galleria Borghese; Bernini and the Amsterdam, Rijksmuseum, SK-A-4233. The woman in the painting of
Birth of Baroque Portrait Sculpture, cat. 1.1. Bernini, self-portrait, c. 1625; the family received two scores in the 'no decision' range (0.652 and
Oxford, Ashmolean Museum; Bernini and the Birth of Baroque Portrait 0.657) when tested against two other known portraits of Eva Visscher.
Sculpture, cat. 3.1. It is possible that the 'no decision' range score in the first pair (Interior
with Eva Visscher, by Van Musscher, c. 1683; Netherlands, private col­
19 The Ottawa Urban VIII, by Bernini; see note 13. Pope Urban VIII, by
lection) is because of the profile in the second portrait (this same private­
Bernini, c. 1631-1632; Rome, Palazzo Barberini; Bernini and the Birth of
ly owned portrait received scores in the non-match range against the
Baroque Portrait Sculpture, cat. 2.4.
Amsterdam and the Cracow likenesses, probably for the same reason;
20 Battista Sforza, by Francesco Laurana, c. 1474; Museo Nazionale del 0.63 and 0.602). In the second portrait of the second score in the 'no de­
Bargello, Florence. C. Damianaki, The Female Portrait Busts of Fran­ cision' range (Eva Visscher, by Van Musscher, c. 1690; Cracow, Wawel
cesco Laurana, Rome, 2000, pp. 55-60. Royal Castle, lnv. 1126), the figure is painted with noticeably larger eyes.

21 Battista Sforza(?), by Francesco Laurana, c. 1472; Musee du Louvre, RF 26 Galileo Ga/ilei (?), artist unknown, c. 1590; Trieste, private collection.
1171. On the cast and its association with the bust of Battista Sforza, see
27 Galileo Galilei (0.724, match range), engraving by Giuseppe Calendi
Damianaki, The Female Portrait Busts, pp. 61-63.
after a lost painting of 1601 by Santi di Tito and published 1793 in Clem­
22 The Piero della Francesca portrait of Battista and the cast received ente de Nelli, Vita e commercia letterario di Galileo Galilei matematico e
a similarity score of 0.664. While this is in the 'no decision' range, it is filosofo. Galileo Galilei (0.718, match range), by Domenico Robusti (Do­
a fairly high score considering that the della Francesca portrait is a pro­ menico Tintoretto), c. 1604; Greenwich, National Maritime Museum, INV
file, something that FACES is not yet able to deal with fully. BHC 2699. Galileo Galilei (0.708, match range), attributed to Filippo di
The portrait busts of Brunelleschi (d. 1446) by Andrea di Lazzaro Nicola Furini (called Lo Sciamerone), c. 1612; Vienna, Kunsthistorisches
Cavlacanti, called Buggiano (Florence, Museo del opera del Duomo), Museum, INV GG 7976. Galileo Galilei (0.672, 'no decision' range), by
and Giovanni Bandini (Florence, Duomo) received a strong score of Ottavio Leoni, 1624; Florence, Biblioteca Marucelliana, lnv. disegni H18.
0.713 when tested against each other. However, Buggiano's 1447-1448 Galileo Galilei (0.65, 'no decision' range), by Domenico Cresti (called
portrait also got a score (0.67) in the 'no decision' range when tested II Passignano), 1624; Helsinki, private collection. Ga/i/eo Galilei (0.639,
against the 1446 death mask cast of Brunelleschi (Florence, Duomo), non-match range), by Justus Sustermans (Suttermans), 1635; Florence,
while Bandini's late sixteenth-century rendering (sometimes said to be Galleria degli Uffizi, lnv. 1890 n. 745. Gali/eo Galilei (0.628, non-match
idealized) obtained a score (0.702) solidly in the match range. This sug­ range), by Charles Mellin, 1639; private collection. Galileo Galilei (0.59,
gests the possibility that Bandini consulted the cast over a century after non-match range), by Justus Sustermans (Suttermans) (?), 1640; Flor­
Brunelleschi's death. ence, Palazzo Pitti, Galleria Palatina lnv. 1912, n. 106.
Somewhat differently, the portrait of Lorenzo de' Medici of around
28 This is said in reference to the Trieste and Vienna portraits.
1553 from the workshop of Bronzino (Florence, Galleria degli Uffizi;
Langedijk, The Portraits of the Medici, cat. 74:6; Bronzino: Artist and 2 9 To cite a range of results, we received one score of 0.71 in the match
Poet at the Court of the Medici, ed. by C. Falciani and A. Natali, Flor­ range for a pair of portraits of Isaac Newton with a probable age differ­
ence, 2010, cat. 11-17), which looks as if it could have been a rendering ence of up to twenty-nine years (though scores in the broader Newton
of the previously mentioned National Gallery of Art bust (as has been paradigm naturally varied from portrait to portrait because, in part, of
suggested, Luchs, 'Lorenzo from Life?', p. 9; probably 1478-1521; artistic quality) (Isaac Newton, by Godfrey Kneller, 1689; Farleigh Wal­
Fig. 8), received a strong match score of 0.7226 against the latter, tend­ lop, Basingstoke, Hampshire; M. Keynes, The Iconography of Sir Isaac
ing to uphold this view. And Vasari's portrait of Lorenzo of 1533-1534 Newton to 1800, Woodbridge, 2005, cat. ll-1 1saac Newton, by Thomas
.

(Florence, Galleria degli Uffizi; Langedijk, The Portraits of the Medici, Murray, 1718; Trinity College, Cambridge; Keynes, The Iconography,
cat. 74:10) - said by the artist to have been modeled on an earlier por­ cat. VII). And for portraits by a great master of a subject he knew well,
trait he felt closely resembled Lorenzo - returned a score in the match we tested two self-portraits by Bernini of around forty-two to forty-sev­
range of 0.7228 when tested against the later portrait by Bronzino, sug­ en years apart, receiving an understandable score in the 'no decision'
gesting that Vasari turned to a likeness that shared a basic body of facial range of 0.684 (Bernini, self-portrait, c. 1623; Rome, Galleria Borghese;
features with the other likenesses of Lorenzo. (The portrait of Lorenzo Bernini and the Birth of Baroque Portrait Sculpture, cat. 1.1. Bernini, self­
by Vasari got a score (0.6747) in the 'no decision' range against the portrait, c. 1665-1670; Windsdr Castle, Royal Library, RL 5539; Bernini
National Gallery bust, such a score probably resulting from the extreme and the Birth of Baroque Portrait Sculpture, cat. 3.13), while those of
angle pose - profile - of the Vasari portrait.) only two years apart got an extremely high score of 0.7342 (Bernini,

289
Conrad Rudolph, Amit Roy-Chowdhury, Ramya Srinivasan, and Jeanette Kohl

self-portrait, c. 1623; Rome, Galleria Borghese; Bernini and the Birth Face. We tested six prominent portraits against it said to be of Shake­
of Baroque Portrait Sculpture, cat. 1 .1. Bernini, self-portrait, c. 1625; speare, receiving scores in the non-match range for five and one in the
Oxford, Ashmolean Museum; Bernini and the Birth of Baroque Portrait 'no decision' range. The scores in the non-match range were with the
Sculpture, cat. 3.1). The suggested crypto-portrait of Michelangelo in Stratford bust (0.6322), the Droeshout engraving (0.6289), the Janssen
the figure of Nicodemus in Caravaggio's Entombment received a score portrait (0.5931), the Cobbe portrait (0.601), and an X-ray of the Hamp­
in the 'no decision' range of 0.655 when tested against Daniele da Vol­ ton Court Palace portrait (0.582). The 'no decision' range score (0.6522)
terra's portrait of Michelangelo, possibly explained by the difference in was with the Chandos portrait.
age (The Entombment of Christ, by Caravaggio, 1602-1604; Vatican
39 For this verse, see Cooper, Searching for Shakespeare, p. 48.
Museums. Michelangelo, by Daniele da Volterra, 1548-1553; Haarlem,
Teylers Museum. On this association, see R. Preimesberger, Paragons 40 Cooper, Searching for Shakespeare, pp. 41-42, 48.
and Paragone, Los Angeles, 2011, pp. 1 oo-107).
41 Cooper, Searching for Shakespeare, pp. 9, 42. On the Chandos portrait,
30 Young Man Among Roses, by Nicholas Hilliard, c. 1587; Victoria and Al­ see Strong, Tudor and Jacobean Portraits, vol. I, pp. 279-283; Cooper,
bert Museum, P.163-1910. R. Strong, Artists of the Tudor Court: The Por­ Searching for Shakespeare, cat. 3.
trait Miniature Rediscovered 152G-1620, London, 1983, p. 9. See also
42 Janssen portrait, artist unknown, c. 1610; Washington, Folger Shake­
R. Strong, Tudor and Jacobean Portraits, 2 vols, London, 1969, vol. I,
speare Library, FPS 17. On the close association of the Stratford bust,
pp. 15-117; Strong, Artists of the Tudor Court, cat. 263; R. Strong,
Droeshout engraving, and Chandos portrait, see Strong, Tudor and Jac­
The Tudor and Stuart Monarchy: Pageantry, Painting, Iconography, 3 vols,
obean Portraits, vol. I, pp. 81, 284. On the Janssen portrait, see Cooper,
Woodbridge, 1995-1998, vol. II, p. 219; L. Campbell et a/., Renaissance
Searching for Shakespeare, pp. 9, 42, cat. 6.
Faces: Van Eyck to Titian, London, 2008, cat. 40.
43 Cobbe portrait, artist unknown, c. 1610; Hatchlands Park, Surrey.
31 Robert Devereux, Second Earl of Essex, by William Segar, 1590; Dublin,
National Gallery of Ireland, NGI.283. Greer, in Campbell, Renaissance 44 That is, 'no decision' range scores with the Cobbe portrait (0.662),
Faces, cat. 40. And see Strong, Artists of the Tudor Court, cat. 218. the Hampton Court portrait (0.673), and William Herbert, Third Earl of
32 Robert Devereux, Second Earl of Essex, by Nicholas Hilliard, 1588; New Pembroke (0.67; by Paul Van Somer, 1617; Royal Collection; RCIN
405870).
York, Metropolitan Museum of Art, 35.89.4. G. Reynolds, European Min­
iatures in the Metropolitan Museum of Art, New York, 1996, cat. 8. 45 Thomas Overbury, Oxford, Bodleian Library, LP 74. On the portraiture of

33 Possibly Robert Devereux, Second Earl of Essex, by Nicholas Hilliard, Overbury, see Strong, Tudor and Jacobean Portraits, vol. I, pp. 238-239.

c. 1588-1590; Edinburgh, National Galleries of Scotland, PG 3423. 46 Cooper, Searching for Shakespeare, p. 68.
Strong, The Tudor and Stuart Monarchy, vol. II, pp. 184-185.
47 See also Strong, Tudor and Jacobean Portraits, vol. I, p. 285, with further
34 The portraits with beards are as follows. Robert Devereux (0.57), by
bibliography.
Isaac Oliver, after 1596; London, National Portrait Gallery, NPG 4966.
48 3 December 2011 to 8 July 2012. See the accompanying publication,
Robert Devereux (0.61), by Marcus Gheeraerts the Younger, c. 1597;
London, National Portrait Gallery, NPG 4985. Robert Devereux (0.57), Imagined Lives.

after Marcus Gheeraerts the Younger, c. 1596; London, National Portrait 49 Artist unknown, c. 1570; National Portrait Gallery, London, NPG 96. On
Gallery, NPG 180. Robert Devereux (0.58), by Isaac Oliver, after 1596; the portraiture of Mary Queen of Scots in general, see Strong, Tudor
London, Victoria and Albert Museum, E.1177-1988. Robert Devereux and Jacobean Portraits, vol. I, pp. 213-223. On this portrait in particu­
(0.6), by Marcus Gheeraerts the Younger, 1596-1601; Greenwich, Na­
lar, Strong, Tudor and Jacobean Portraits, vol. I, pp. 217-218; Imagined
tional Maritime Museum, BHC2681. Robert Devereux (0.62), by Isaac Ol­
Lives, p. 92.
iver, 1596-1598; Royal Collection, RCIN 420933. Robert Devereux (0.61),
50 The scores in match range are as follows. Mary Queen of Scots (0. 7), by
artist unknown, c. 1599; Cambridge, Trinity College, TC Oils P 53. Rob­
ert Devereux (0.62), by Marcus Gheeraerts the Younger, c. 1596-1599; Franc;ois Clouet, c. 1558; Royal Collection, RCIN 401229. Mary Queen
Cambridge, Trinity College; TC Oils P 52. Robert Devereux (0.63), Mar­ of Scots (0.7112), by Franc;ois Clouet, c. 156G-1561; Royal Collection,
cus Gheeraerts the Younger, c. 1596; the Marquess of Tavistock and the RCIN 403429. Mary Queen of Scots (0.72), by Franc;ois Clouet and
Trustees of Bedford Settled Estates, Woburn Abbey. workshop, c. 1555; Wroctaw, Zaktad Narodowy im. Ossolinskich. Mary
Without beard: Robert Devereux (0.63), attributed to the studio of Queen of Scots (0.7), by Nicholas Hilliard, 1579; Royal Collection, RCIN
Nicholas Hilliard, c. 1587; London, National Portrait Gallery, NPG 6241. 420641.
The 'no decision' scores are as follows. Mary Queen of Scots
35 'The Moost Happi' medal prototype 1534; London, British Museum.
(0.6842), by Franc;ois Clouet or Jacques Decourt, c. 1558; Paris, Biblio­
36 In general, see Strong, Tudor and Jacobean Portraits, vol. I, pp. 276- theque nationale de France, Estampes, Res. Na 22. Mary Queen of Scots
286; T. Cooper, Searching for Shakespeare, New Haven, 2006. (0.6873), by Franc;ois Clouet, c. 1560; Paris, Bibliotheque nationale de
France, Estampes, Res. Na 22. Mary Queen of Scots (0.6743), artist
37 Cooper, Searching for Shakespeare, pp. 47-51. On the Stratford bust,
unknown, c. 156G-1592; London, National Portrait Gallery, NPG 1766.
see Strong, Tudor and Jacobean Portraits, vol. I, pp. 283-286; Cooper,
Mary Stuart (and Franc;ois II of France) (0.69), workshop of Franc;ois
Searching for Shakespeare, cat. 2.
Clouet, c. 1558; Paris, Bibliotheque nationale de France, ms NAL 82.
38 On the Droeshout engraving, Strong, Tudor and Jacobean Portraits,
51 James Scott, by Abraham Blooteling, after Sir Peter Lely, c. 166G-1690;
vol. I, p. 283; Cooper, Searching for Shakespeare, cat. 1. It has been said
National Portrait Gallery, London, NPG D19810.
by some that the so-called 'Kesselstadt Death Mask' (inscribed 1616;
Darmstadt, Universitats- und Landesbibliothek) is a cast from a death 52 For example, the clear visual match between Bernini's bust of Fran­
mask of Shakespeare himself; Hammerschmidt-Hummel, The True cesco I d'Este (165G-1651; Galleria Estense, Modena) and the exacting

290
FACES: Faces, Art, and Computerized Evaluation Systems- A Feasibility Study of the Application of Face Recognition Technology...

representation of this same bust in the late seventeenth-century painting 55 For a paradigm that exemplifies this issue, with two likenesses that ap­
attributed to Francesco Stringa (Minneapolis, Minneapolis Institute of pear to be identical to the naked eye - one of terracotta and the other of
Arts; discussed inS. Ostrow, '(Re)presenting Francesco I d'Este: AnAl­ marble - but that receives a non-match score (0.6204), ct. Muzio Fran­
legoricalStill Life in the Minneapolis Institute ofArts', Artibus et Historiae, gipane, by Algardi, probably c. 1635, terracotta, Bologna, Pinacoteca
63, 2011, pp. 201-216) received a 'no decision' range score of 0.6824 (J. Montagu, Alessandro Algardi, New Haven, 1985, val. II, fig. 150); and
apparently because of the limiting of facial information resulting from the Muzio Frangipane, by Algardi, probably c. 1635, marble; Rome, S. Mar­
rendering in profile. cello al Corso, Capella Frangipane (Montagu, val. II, fig. 151). And for
an interesting variation on this problem, the portrait sometimes said to
53 On the draft of a proclamation of 1563 (not published at the time) that,
be of Mary Queen ofScots discussed above (Fig. 20; London, National
in part, would have attempted to limit unsuitable images of Queen
Elizabeth, see T. Cooper, The Queen's Visual Presence', in Elizabeth: Portrait Gallery, NPG 96) received a score in the match range (0.7112)

The Exhibition at the National Maritime Museum, ed. by S. Doran, Lon­ when tested against a known oil painting of Mary by Clouet (Royal Col­

don, 2003, p. 178. Flattery, of course, was not expected to be limited to lection, RCIN 403429) but got a 'no decision' range score (0.6873) with

queens; for Mantegna's perceived deficiency in this regard, see Chris­ an otherwise visually 'identical' chalk study by Clouet for the painting
tiansen and Weppelmann, The Renaissance Portrait, pp. 5-6. (Paris, Bibliotheque nationale de France, Res. Na 22).

54 This seems to be the cause of the 'no decision' range score with the 56 On this, see C. B. Donnan, Moche Portraits from Ancient Peru, Austin,
Monmouth engraving mentioned above. 2004, pp. 141-159. Our thanks and regrets to Professor Donnan.

291

Você também pode gostar