US 20140185959A1

(19) United States

(12) Patent Application Publication (10) Pub. No.: US 2014/0185959 A1
(43) Pub. Date:

Coppin et al.
(54)

METHOD AND APPARATUS FOR IMAGE

Jul. 3, 2014

Publication Classi?cation

PROCESSING

(51)

(71) Applicant: DeepMind Technologies Limited,

London (GB)

Int. Cl.
G06F 1 7/30

(2006.01)

(52) U.S.Cl.
CPC .............................. .. G06F17/30262 (2013.01)

(72) Inventors: Benjamin Kenneth Coppin, Cottenham

(GB); Mustafa Suleyman, London
(GB); Arun Nair, London (GB)

(73) Assignee: DeepMind Technologies Limited,

London (GB)

USPC

(57)

........................................................ ..

382/305

ABSTRACT

A method for processing an image to generate a signature

Which is characteristic of a pattern Within said image. The

method comprising receiving an image; overlaying a Window

(21) Appl.N0.: 14/155,434

at multiple locations on said image to de?ne a plurality of

(22) Filed:

sub-images Within said image, With each sub-image each

having a plurality of pixels having a luminance level; deter
mining a luminance value for each said sub-image, Wherein

Jan. 15, 2014

Related US. Application Data

(63)

(30)

Continuation of application No. 13/ 801,027, ?led on

Mar. 13, 2013, now Pat. No. 8,644,607.

Foreign Application Priority Data

said luminance value is derived from said luminance levels of

said plurality of pixels; and combining said luminance values

for each of said sub-images to form said signature. Said
combining is such that said signature is independent of the
location of each sub-image. A method of creating a database

of images using said method of generating signatures is also

Nov. 1, 2012

(GB) ................................. .. 1219650.7

described.

Patent Application Publication

Jul. 3, 2014

Sheet 1 0f 11

Fig '11::

Fig '32

Fig w

Fig 1i;

Fig 1e

US 2014/0185959 A1

Patent Application Publication

Jul. 3, 2014

Sheet 2 0f 11

Receive image
83 CG

Canve? to gzeyscale (optima?)

8102

Rascals t0 a ?xeci siza

$184

De?ne hisiogram
8186
if

Pepuiate histagram based on

rescaled image
3188
if

Comes? histagram in a vector or

artth signature far the image

31m

Fig 2a

US 2014/0185959 A1

Patent Application Publication

Jul. 3, 2014

Sheet 3 0f 11

US 2014/0185959 A1

{Jaime a wincmw

8286
v

Overlay wincow on image

8202

%
Reduce image within overlaid Window to
a ?xed number a? pixes
S204

%
{Define threshold
$285
{38mm average iuminame

for each pixel

8206

ihan ?wreshoid

a Set to black 5

\\ 328%,.f/f

Satin white
$212

32113
. . . . . . . . . . . . . . .

Convert ta binary
8214
............................. ._

Ganvezi 1:: index numbar

Fig 2b

8216
a?

Bncremwt siui in histugsam with maici?ng

index numhes"
SZi 8

Patent Application Publication

Jul. 3, 2014

Sheet 4 0f 11

US 2014/0185959 A1

Fig3c
Fig
5

Fig.3i
3a
Fig

Patent Application Publication

Jul. 3, 2014

Sheet 5 0f 11

Hnowizms

H$2m5a.0
mmHoHvlo

cow

do?

bow

on

89

com

com

bow

US 2014/0185959 A1

Patent Application Publication

Jul. 3, 2014

Sheet 6 0f 11

Lisas

{Computer
160

i522.

\f

gewer
Opgfg??g
Syatem
/\

m
Application
Programs

{rogram
maze
1\

My
\MW

image
database

fieatufe
database

3.2%

iiiii

US 2014/0185959 A1

Patent Application Publication

v?

Jul. 3, 2014

Sheet 7 0f 11

US 2014/0185959 A1

Patent Application Publicat 10]]

Jul. 3, 2014

Sheet 8 0f 11

US 2014/0185959 A1

Patent Application Publication

Jul. 3, 2014

Sheet 9 0f 11

Receive image for szeamh queyy

83GB

Get twistogzam for

raceived image
5362
v

Retrieve histmgram for

amiher image
3384
V

Determine a similarity value far

{we images
833%

ql

<

\\

?} ......................................................... ._

yes

Rank similarity values for each

pair of images
S322

Qutpu highest ranked images

3314

Fig "Fe

US 2014/0185959 A1

Patent Application Publication

Jul. 3, 2014

Fig 88

Sheet 10 0f 11

Fig; 82';

Fig 8c

De?ne ramesem'a?ve @a?em for

aaah sub~sectinn

84M
V

Determine: signatme far

remeseniat'ive gaitem
S492

US 2014/0185959 A1

Patent Application Publication

Jul. 3, 2014

Sheet 11 0f 11

Receive image far cassi?sation

$580

Determine signaiure for

received image-2
$562
V

Qetermine a simiariiy value

hammer} QLEQW signatura and

each Eibrary sub~seeiion
S534

Rank the similariiy values

850%
F

Assign the label far the ciasest

rankeii sub~seatians>
55%

{Omiona} Determine damnant

comm" Qf ques'y image
S51 8

Cambna labs? assigned wiih

label far dominant aoaur
$.35? 2

Fig 8e

US 2014/0185959 A1

Jul. 3, 2014

US 2014/0185959 A1

METHOD AND APPARATUS FOR IMAGE

PROCESSING
CROSS REFERENCE TO RELATED
APPLICATION

[0001]

The present application is a continuation of US. Pat.

No. 13/801,027 ?led Mar. 13, 2013; which itself claims pri
ority to British Patent Application No. GB 12196507 ?led
Nov. 1 , 201 2. The entirety of the aforementioned references is

incorporated herein by reference for all purposes.

TECHNICAL FIELD

[0002] This invention relates to image processing, in par

ticular to processing images to assist in image searching and

image matching.
BACKGROUND OF THE INVENTION

between different types of rock from close-up photos. How

ever, these methods do not work well with all images, in
particular, with images such as those shown in FIGS. 1a to 1e
which have a coarse grain and/or comprise more than one

texture within the image. The applicant has thus recognized

the need for an improved method and apparatus for process

ing images to improve image matching and/ or searching.

[0012] Another method for matching images is described in
U.S.2009324077 to Microsoft Corporation entitled Patch

Based Texture Histogram Coding for Fast Image Similarity

Search. As set out in the abstract, this describes techniques

for identifying copied images based on an original image.

Identifying a copied image is based on creating unique and
identi?able features that in turn are used to generate multiple

histograms. The histograms are generated by patches of the

image, where the patches are created by equally dividing the
image. The combined patch histograms are representative of
the image.

[0003] FIGS. 1a to 1e show a variety of static two-dimen

sional images, in this case photographs of clothing or home

SUMMARY OF THE INVENTION

furnishing. Each image contains many elements, including

color (black, white, grey, red, orange), an object (e.g. dress,
shorts, shirt or rug) and texture (e.g. stripes, spots, checks). In

provided a method for processing an image to generate a

signature which is characteristic of a pattern within said

each image, the texture has a large scale relative to the image

image comprising:

[0013]

According to a ?rst aspect of the invention, there is

itself. Thus the texture may be termed macro-scale or coarse

[0014] receiving an image;

grained. Moreover, the texture is not consistent across the

entire image. The images of FIGS. 1a to 1e are merely rep

[0015]

resentative of the wide variety of images which fall into this

category. By contrast, some images contain just a single tex
ture and that texture is ?ne-grained, e. g. the grain pattern in a

rock. The word pattern may be used interchangeably with

texture.

[0004]

It would be useful to be able to classify and/or search

for images based on their texture or pattern. There are various

known methods for classifying or searching by texture,

namely statistical models, geometrical methods, model

based methods and signal-processing methods.

overlaying a window at multiple locations on said

image to de?ne a plurality of sub-images within said image,

with each sub-image each having a plurality of pixels having
a luminance level;
[0016] determining a luminance value for each said sub
image, wherein said luminance value is derived from said

luminance levels of said plurality of pixels;

[0017] combining said luminance values for each of said
sub-images to form said signature;
[0018] wherein said combining is such that said signature is
independent of the location of each sub-image.
[0019] By creating a signature that is independent of the
location of each sub-image, this method is insensitive to

Statistical Models

[0005] Typically, these methods assume that there is just

one texture in the entire image. Alternatively they can be

applied to parts of images to identify texture in different parts

of the image. One example of such a technique is the use of
co-occurrence matrices described in Robert Haralicks Tex

tural Features for Image Classi?cation published in IEEE

Transactions on Systems, Man and Cybernetics in November
1 973.

[0006]

Geometrical Methods

[0007]

These methods assume that images are made up of

atomic textural units which can be identi?ed. This can work

well with very regular patterns.

[0008] Model-Based Methods
[0009]

These use mathematical models to build up the com

ponents of the texture in an image.

location (i.e. it is translation independent). In other words, it

will identify patterns wherever they occur in an image.
Hence, an image that contains a stripy dress in the center
could have a similar signature to an image in which the stripes
are around the border of the image.
[0020] This invention provides a way of comparing the
gross-scale texture or pattern of one image with that of a
number of other images. This can be used, for example, to
search for similar images based on pattern.
[0021] Said luminance value may be determined by deter
mining an average luminance for each of said plurality of

pixels; determining whether said average luminance is

greater than a threshold value; setting each of said plurality of
pixels to a ?rst value if said average luminance is greater than
a threshold value; setting each of said plurality of pixels to a
second value if said average luminance is lower than a thresh

old value; and combining said ?rst values and said second

Signal Processing Methods

values to generate said luminance value for each said sub

image. Said threshold value may be the average luminance of

[0010]

the image.

This involves using methods such as Fourier trans

forms and wavelets, effectively looking at the frequencies (of

changes in color or luminance) contained within the image.

[0022] Said ?rst value may be black or one and said second
value may be white or zero. The use of black and white results

[0011]

in a simple sub-image composed of black and white pixels.

These existing methods for matching texture typi

cally work well with images that contain just a single texture
and where the texture is ?ne-grained. This is useful, for

example, for automatically determining the difference

The black and white pixels may be converted into a luminance

value which is a binary number. The binary number may be
further converted into an index number. It will be appreciated

Jul. 3, 2014

US 2014/0185959 A1

that the binary and index numbers may be determined in

different ways; all that is essential is that the same method is

used for all sub-images.

[0023] Combining said luminance value may comprise
de?ning a histogram having a plurality of slots with each slot
representing a luminance value. The slots may thus be num
bered with the index number where one is determined. Each

slot in said histogram may be incremented for each sub-image

having a luminance value corresponding to said slot. In this
way, the combining ensures that the signature is independent
of the location of each sub-image because the slot in the
histogram which corresponds to the index number is incre
mented regardless of the location within the image of the
sub-image which led to that index number.
[0024] The size of the histogram may be determined from
the size of each sub-image. As an example, said histogram
may have 512 slots and each said sub-image may have 3x3

pixels.

which is here incorporated by reference. Thus, the features set

out below are designed to be incorporated in the following

aspects.
[0031] According to another aspect of the invention, there
is provided a method of searching for an image comprising:
[0032] receiving a query image;

[0033] determining said signature for said query image,

wherein said signature is generated using the method
described above;
[0034]

retrieving an image from a database together with its

corresponding signature;
[0035]

comparing said signature for said query image with

said signature for said retrieved image to generate a similarity

value for said retrieved image;

[0036] repeating said retrieving and comparing steps to

generate similarity values for multiple retrieved images;
[0037] ranking said similarity values; and
[0038] outputting said retrieved images having highest

[0025] Said signature may be considered to be said histo

gram. Alternatively, said histogram may be converted to a
feature vector to form said signature.
[0026] The use of a histogram means that the method falls

ranks.

into the statistical category. However, unlike existing meth

de?ning a plurality of sub-sections within said database with

each sub-section having a corresponding signature which is

ods, our method works well with macro-scale textures con

tained in real-world images with multiple patterns and tex

tures.

[0039]

According to another aspect of the invention, there

is provided a method of creating a database of images com

prising:
generated using the method described above;
[0040] receiving an input image to be input into said data

[0027] The window may be overlaid on the image such that

at least some of said plurality of images partially overlap at
least one other image. In other words, in contrast to U.S.2009/
0324077, the image is not divided up into separate sub-im
ages. The overlapping images will reinforce the dominant

base;

patterns within the image as a whole and ensure that patterns

are not missed that happen not to span a particular window.
Another key difference over U.S.2009/ 0324077 is that a

similarity value for each of said plurality of sub-sections;

single histogram is calculated for the whole image rather than

a histogram being calculated for each sub-image. Further
more, the present invention calculates a luminance value to

determine the pattern whereas in U.S .2009/ 0324077, textures

are calculated based on dominant directions. The use of domi

nant directions means, for example, that while it may be able

to deal fairly well with stripes, the method of U.S.2009/
0324077 would not work as well with spots, leopard-skin or
other non linear patterns.

[0028]

In the present invention, the image may be rescaled

[0041] determining a signature for said input image using

the method described above;

[0042] comparing said signature for said input image with

each signature of said plurality of sub-sections to generate a

[0043] ranking said similarity values; and

[0044] allocating said input image to said sub-section hav
ing the highest rank.
[0045] The determining of a signature for said query image
and/or said input image may be accessing said signature from
a database. Alternatively, said signature may be calculated in
real-time using the method described above.
[0046] In both of the above aspects, the method may be
combined with other features such as color and shape. In

other words, the method may further comprise inputting a

color and/or shape with the query image and generating a

similarity value which incorporates said color and/or shape.

In other words, the method may further comprise determining
the color and/ or shape of the input image and the input image

before overlaying said window. The rescaling ensures a con

sistent result because all images are rescaled to the same ?xed
size. Said sub-image within said window may also be res

may be allocated to a sub-section depending on its color

caled to reduce the number of pixels within said sub-image to

and/or shape.

said plurality of pixels. Reducing the number of pixels affects

the computation of the signature for the overall image. For

[0047] The invention further provides processor control

code to implement the above-described systems and methods,

example, if too many pixels are contained in each sub-image,

there will be an unmanageable number of slots in the histo
gram.

digital signal processor (DSP). The code is provided on a

physical data carrier such as a disk, CD- or DVD-ROM,

[0029]

programmed memory such as non-volatile memory (eg

Conversion to grayscale may also be incorporated

for example on a general purpose computer system or on a

into said method. The color is not generally relevant to the

Flash) or read-only memory (Firmware). Code (and/or data)

pattern of the image. It is the luminance of each pixel which

is key. For example, said received image or said rescaled
image may be converted to grayscale before determining said
luminance values. Alternatively, the conversion to grayscale
may take place when determining said luminance value for

to implement embodiments of the invention may comprise

source, object or executable code in a conventional program

said sub-image.

ming language (interpreted or compiled) such as C, or assem

bly code. As the skilled person will appreciate such code
and/or data may be distributed between a plurality of coupled
components in communication with one another.

[0030] The method was designed to be incorporated into an

image search or image matching system such as that
described in our pending patent application GB121518.3

out the methods described above. Thus, according to another

aspect of the invention, there is provided a system for pro

[0048]

The invention further provides systems for carrying

Jul. 3, 2014

US 2014/0185959 A1

cessing an image to generate a signature which is character

istic of a pattern within said image, said system comprising
[0049]
[0050]
[0051]

an input to receive an image; and

a processor con?gured to:
overlay a window at multiple locations on said

[0074] allocate said input image to said sub-section having

the highest rank.
[0075] In each of the systems described above, the signa
ture of the query image or input image may be created using
the methods described in detail above.

image to de?ne a plurality of sub-images within said image,

with each sub-image each having a plurality of pixels having

BRIEF DESCRIPTION OF THE DRAWINGS

a luminance level;
[0052] determine a luminance value for each said sub-im
age, wherein said luminance value is derived from said lumi

grained texture;

nance levels of said plurality of pixels;

[0053] combine said luminance values for each of said

processing an image according to the present invention;

sub-images to form said signature;

[0054] wherein said combining is such that said signature is
independent of the location of each sub-image.
[0055] According to another aspect of the invention, there

2a;

is provided a query engine for searching for an image com

prising:
[0056]
[0057]

an input for receiving a query image; and

a processor con?gured to

[0058]

determine a signature for said query image by

[0059]

overlaying a window at multiple locations on said

image to de?ne a plurality of sub-images within said image,

with each sub-image each having a plurality of pixels having
a luminance level,

[0060] determining a luminance value for each said sub

image, wherein said luminance value is derived from said

luminance levels of said plurality of pixels, and combining

said luminance values for each of said sub-images to form

said signature; wherein said combining is such that said sig

nature is independent of the location of each sub-image;
[0061] retrieve an image from a database together with its

corresponding signature;
[0062] compare said signature for said query image with
said signature for said retrieved image to generate a similarity
value for said retrieved image;
[0063] repeat said retrieving and comparing steps to gener
ate similarity values for multiple retrieved images; and

[0064] rank said similarity values; and

[0065] an output for outputting said retrieved images hav

ing highest ranks.

[0066]

According to another aspect of the invention, there

is provided a system for characterizing an image within a

database of images, the system comprising:

[0067]

a database having a plurality of sub-sections with

each sub-section having a corresponding signature;

[0068] an input for receiving an input image to be input into
said database;
[0069]

a processor which is con?gured to determine a sig

nature for said input image by overlaying a window at mul

tiple locations on said image to de?ne a plurality of sub

images within said image, with each sub-image each having a

plurality of pixels having a luminance level;
[0070] determining a luminance value for each said sub
image, wherein said luminance value is derived from said
luminance levels of said plurality of pixels; and
[0071] combining said luminance values for each of said
sub-images to form said signature; wherein said combining is
such that said signature is independent of the location of each

[0076]

[0077]

FIGS. 1a to 1e show various images having a coarse

FIG. 2a is a ?owchart setting out the method for

[0078] FIG. 2b is a ?owchart setting out the steps associ

ated with de?ning a histogram as part of the method of FIG.

[0079]

FIG. 3a illustrates an image to be processed accord

ing to the present invention;

[0080] FIG. 3b illustrates a rescaled version of the image of
FIG. 3a;
[0081] FIG. 30 illustrates a sub-image of image of FIG. 3b;
[0082] FIG. 3d illustrates a processed version of the sub

image of FIG. 3d;

[0083] FIG. 4 illustrates a histogram created using the steps
of FIG. 2b;
[0084] FIG. 5 illustrates the image of FIG. 3b with a win
dow overlaid at various positions;
[0085] FIG. 6 schematically illustrates a system for imple

menting the present invention;

[0086]

FIGS. 7a to 7d illustrate the results for searching for

images which match an image which has been processed

according to the present invention;
[0087] FIG. 7e is a ?owchart of the method of searching
used in FIGS. 7a to 7d;
[0088] FIGS. 8a to 80 show images which are used to build

up a library of images using the processing techniques of the

present invention;
[0089]

FIGS. 8d and 8e are ?owcharts of the method

involved in building up a library of images and categorizing

images.
DETAILED DESCRIPTION OF THE DRAWINGS

[0090] FIGS. 2a and 2b illustrate the steps in the processing

of an image according to the present invention. The overall
result is that the pattern/texture feature of the image is effec

tively extracted from the image. The ?rst step (S100) is that
the image is received.
[0091] At step S102, the image is converted to grey-scale
(as color is not relevant). This is an optional step. Various
techniques for the conversion may be used. One method is to
simply convert each RGB color pixel to its luminance value

using a linear combination of the gamma-compressed pri

mary intensities for each color. For example, using the ITU-R
601 -2 luma transform:
LIR >299/1000+G*587/1000+B*1 14/1000

[0092] where L:luminance; R:red value; GIgreen value;

BIblue value.
[0093] The next step S104 is to rescale the image to a ?xed
size (say 100x100 pixels). This is also illustrated in FIGS. 3a

sub-image;

and 3b. FIG. 3a shows the original image (after optional

grayscale conversion) and FIG. 3b shows the rescaled image.

[0072]

compare said signature for said input image with

The rescaling ensures a consistent result because all images

each signature of said plurality of sub-sections to generate a

are rescaled to the same ?xed size. It also removes some of the

similarity value for each of said plurality of sub-sections;

[0073] rank said similarity values; and

?ne detail from very large images, which is consistent with a

desire to analyze only the coarser grained patterns in the

Jul. 3, 2014

US 2014/0185959 A1

image. Rescaling (or resizing) of images can be done in a

number of ways, for example, nearest neighbor, linear inter
polation in a 2x2 environment, cubic spline interpolation in a
4x4 environment, or down sampling with anti-aliasing.

[0100]

The ?xed number of pixels are ?attened into a

binary number, by converting black pixels to 1 and white

may be used. This algorithm calculates the horizontal and

vertical ratios between the original image and the scaled
image. Once the ratios have been calculated, a buffer is cre
ated to store information for the scaled image which is

pixels to 0 (step S214). This binary number is converted to an

index number (step S216) which corresponds to a slot in the
histogram. The histogram slot with the speci?ed index has its
value incremented by 1 at step S218. For example, the binary
number for the 9 pixel image of FIG. 3d is 111000111 and
thus the corresponding index number is 455. It will be appre
ciated that the binary and index numbers may be determined

derived from the original image. The technique is described in

more detail in Nearest Neighbor Image Scaling published

is used for all images and sub-images. The binary or index

[0094] For example, the nearest neighbor scaling algorithm

in Tech-Algorithm.com on 6 Oct. 2007.

[0095] The next step (S106) is to de?ne a histogram with a

?xed number of slots. In the example below, an empty histo

gram is de?ned with 512 slots (the reason for this number is

explained below). It is preferable to initialize the histogram

before use, e.g. by inserting 0 into each slot. At step S108, the
histogram is populated based on the rescaled image, as
described in more detail in FIG. 3b. Finally, at step S110, the
histogram is converted to a vector of non-negative integer

in different ways; all that is essential is that the same method

number may be considered to be a luminance value which is

representative of the luminance levels within the sub-image.

It is also noted that the binary and index numbers are location

independent. The slot in the histogram which corresponds to

the index number is incremented regardless of the location
within the image of the sub-image which led to that index
number. This is a key feature of the invention.
[0101] The next step S220 is to determine whether or not
the window can be moved to overlay a new block of the

values. The vector has the same number of values as there are

image. As illustrated in FIG. 5, after processing a chunk of the

slots in the histogram. This vector can be thought of as a

image falling within the window, the window is moved

signature which is characteristic of the pattern within the

image or alternatively the histogram itself may be considered
to be the signature. Other techniques may be used for creating
the pattem-signature if desired. However, as explained below,
the signature is a translation-independent representation of
the pattem(s) contained in the image.
[0096] FIG. 2b explains in more detail the steps in populat
ing the histogram based on the rescaled image. In the initial

around the image. The window covers every sub-image of a

step S200 a window is de?ned on the rescaled image. Such a

window is shown in FIG. 5 as a square window which is small

relative to the image on which it is placed. Thus, for an image

rescaled to 100x100 pixels, a window of 12x12 pixels may be
used. It will be appreciated that other shapes of window may
be used although a square simpli?es the calculations.
[0097] At step S202, the window is overlaid on the image to
create a sub-image of the image. In the present example, the
sub-image has 12x12 pixels and is illustrated in FIG. 30. The
12><12 sub-image is shrunk to 3x3 pixels at step S204. As
explained below, a larger ?xed number could be used but the
ef?ciency of the algorithm decreases rapidly as this number
increases. A threshold is de?ned at step S205. This threshold
value may be the average luminance of the rescaled image and
may be calculated using standard techniques. The average
luminance value for the 9 pixels is calculated at step S206.
This average luminance value is calculated in any standard
way.
[0098]

At step S208, it is determined whether or not the

luminance of each pixel is greater than the threshold value. If
the luminance value is greater than the threshold value, the

pixels are set to black (S210), and all others are set to white

(S212). This results in an image containing only pixels which

are wholly black or wholly white pixels as illustrated in FIG.
3d.

[0099] The thresholding step described above means that

the grayscale conversion is not essential. This is possible
because the thresholding step could be carried out on the
luminance of RGB colored cells. In other words, each RGB
value could be converted to a luminance value in order to

calculate the average luminance for each pixel at step S206.

The thresholding could then be applied at Step S208 to this
luminance value.

?xed size (say 12x12) in the image. In our experiments, we

used a window of 12x12 (i.e. 144 pixels in total) and moved
it by one pixel at each step. The window moves across from

the top-left to the top-right with each step. Following this, the
window would move down by one pixel and would start again

from the left-hand size moving to the right. This would be

repeated until the entire image had been covered. This move
ment is illustrated in FIG. 5 with the size of the window and

the distance moved by the window exaggerated for clarity. As

shown in FIG. 5, both the ?rst and second placements of the
window capture the left strap of the dress. Similarly, both the
fourth and ?fth placements of the window capture the right
strap of the dress. Thus, there is overlap between the sub
images captured by the window. Such overlap is a preferred
feature of the present invention.
[0102] Once the window has been moved to a new location,
steps S204 to S220 are repeated for the new location. Thus,
the method is an iterative process which gradually covers
every sub-image. Once there are no more sub-images to pro

cess, the histogram is output.

[0103] An example ofa histogram is shown in FIG. 4. The
histogram that is created contains a representation of the
types of patterns that are present in the image, in this case for
the dress of FIG. 3a. The y-axis shows the count in each slot
in the histogram, and the x-axis shows the slot numbers. The

dominant patterns are represented by larger values than pat

terns that occur in only small parts of the image. In this case,
the peaks in the histogram are at 0, 56, 455 and 504. The nine
pixels sub-images for the four non-zero peaks are shown as
inserts. Three of these four dominant patterns represent the
horizontal stripes of a certain thickness.Another smaller peak
occurs at 219 which represents the edge of each dark hori

zontal stripe.
[0104] 0 is the most dominant pattern. This is mostly
because of the background being white, which counts as
000000000. This is not necessarily a problem but the method
can be improved by applying automatic segmentation to sepa
rate the main image from the background and then the method

described here applied only to the foreground. This would, in

this case, eliminate the vast majority of the values in the 0 slot,

Jul. 3, 2014

US 2014/0185959 A1

although 0 would still probably be a minor peak because of

the horizontal white stripes within the main image.
[0105] An image with similar patterns will have a similar
histogram, whereas an image with very different patterns will
have a very different histogram overall.

[01 06]

The numbers used in the descriptions here have been

selected as optimal based on experimentation. However,

other values can be used. The size of the histogram is entirely
dependent on the size that each sub-image is resized to. In this
case, the minimum ?xed size used is 3x3 pixels. The smallest

binary number possible is 000000000 (which in decimal is

zero) and the largest possible, with nine bits, is lllllllll
which in decimal notation is 51 1. Hence, the number of slots

in the histogram in this example is 512.

[0107] In general, the number of slots is de?ned by:

[0108] 11:21
[0109]

Where d><d is the number of pixels in the ?nal ver

sion of each sub-image (in our case, 3x3:9). Using larger

sub-images results in signi?cantly larger histograms. For

example, having 4><4 sub-images leads to 216:65,536 slots.

(I/O) Controller, a system power and clock source; display

driver; RAM; ROM; and a hard disk drive. A network inter
face provides connection to a computer network such as Eth

ernet, TCP/IP or other popular protocol network interfaces.

The functionality may be embodied in software residing in
computer-readable media (such as the hard drive, RAM, or
ROM). A typical software hierarchy for the system can
include a BIOS (Basic Input Output System) which is a set of
low level computer hardware instructions, usually stored in
ROM, for communications between an operating system,
device driver(s) and hardware. Device drivers are hardware
speci?c code used to communicate between the operating
system and hardware peripherals. Applications are software
applications written typically in C/C++, Java, assembler or
equivalent which implement the desired functionality, run
ning on top of and thus dependent on the operating system for
interaction with other software code and hardware. The oper
ating system loads after BIOS initializes, and controls and
runs the hardware. Examples of operating systems include
LinuxTM, SolarisTM, UnixTM, OSXTM Windows XPTM and

5x5Q22SIover 33 million slots.

[0110] FIG. 6 shows a schematic image of a system which

equivalents.

can be used to implement the various processes described

herein. The system can be used in many ways, including
allowing a user to make searches and categorizing images in
an image database.
[0111] Where a user is inputting data, e.g. an image to be

a query image. In each Figure, the ?rst image is the query

image and the following images are the ones deemed by a
searching algorithm to have the most similar pattern/texture.

searched or categorized, a user can use an input device 100.

The input device is shown as a computer but may be any

suitable device, including computers, laptops, mobile

phones. Each input device typically comprises one or more

[0115]

FIGS. 7a to 7d show the results of a search based on

The searching algorithm is set out in FIG. 7e as follows:

[0116] The ?rst step S300 is to receive an image for a search
query. This may be input by a user by any known method. The

next step S302 is to get the histogram for the received image.
If the query image has been processed using the method
described above, the histogram or the corresponding pattern

processors, memory, a display, a user interface, e.g. devices

feature vector may be retrieved from a database. Altema

such as keypad, keyboard, microphone, touch screen and a

network interface. The input device 100 is connected in any

tively, the histogram may be generated in real-time as part of

known manner, e.g. over the Internet 102, to a server 104 for

the query.
[0117] The next step S304 is to retrieve a histogram or

processing any user input.

corresponding pattern feature vector for another image. The

[0112] The server 104 comprises a processor or operating

system which processing any data received at the server. The

image and its associated histogram are selected from a data

base, which may have over 1,000,000 images. This database

server 104 also comprises memory storing applications pro

may comprise pattern feature vectors for all images in the

grams and program data. These application programs are run

database, pre-computed and stored.

to create the desired functionality. For example, if the server

104 receives an image from a user, the server may run the

[0118] The two histograms can now be compared using

standard techniques, e.g. earth-mover distance, Euclidean

application program which generates the signature for this

distance or another vector or histogram difference metric.

image as described in FIGS. 2a and 2b. Once the signature has

been generated, this may be used to search for comparable

the patterns contained in the two images. The difference can

images as described in more detail in relation to FIG. 7e or to

categorize the image for storage in a library as described in

relation to FIG. 8d.
[0113] As shown in FIG. 6, the server is connected to an
image database 106 and a feature database 108. The feature
database 108 may store the signatures created as described

above. The image database 106 may store the image associ
ated with these signatures. When a user is selecting an image
for searching, an image(s) from the image database may be
presented to a user on the user interface of the input device

100 for selection. When searching for comparable images, the

server may query the images and associated signatures in
these databases. The image database 106 and a feature data

Such techniques typically determine the difference between

be converted to a similarity by inverting it (for example by

subtracting it from the maximum possible difference score).
A similarity value for the pair of images is thus determined
S308.
[0119] The process S310 then determines whether or not
there are any other images to be compared in the database. If
there are other images, steps S304 and S308 are repeated.

Thus, multiple similarity values are generated in which the

query image is compared with other images. These similarity
values can then be ranked S312 and the highest ranking
images S314 may be output. In other words, a k-nearest

neighbor algorithm is applied to identify the nearest neigh

base 108 may also be termed a library as described in more

bors to the query image, which in this case will be the images
with the most similar patterns to those contained in the query

detail below.

image.

[0114]

[0120] In FIG. 7a, the dominant pattern of the query image

is made up, approximately, of horizontal stripes. The algo

The server may be implemented using standard

technology as a single unit or a plurality of interconnected

units. The hardware components of any server typically

include: a central processing unit (CPU), an Input/Output

rithm has correctly identi?ed, as the ?ve most similar images,

other items of clothing with bumpy horizontal stripes. Note

Jul. 3, 2014

US 2014/0185959 A1

that color and shape have not played a part; the matching has
been done solely on the patterns found in the query image.
[0121] The query image in FIG. 7b is almost entirely made
up of diagonal stripes. The ?rst two results identi?ed by the

[0131] iii. Compare the query images histogram with that

of each dictionary pattern in turn, determining the distance
(e.g. Euclidean distance) between the two histograms (step

S504).

algorithm are similarly dominated by diagonal stripes. The

[0132]

next three results can also be seen to have diagonal patterns

[0133] v. Select the pattern whose histogram is closest to

that of the query image and look up the textual description of

working from the bottom-left to the top-right, although the

pattern is less visually obvious than in the query image.
[0122] In the third query image of FIG. 70, the pattern is
made up of approximately vertical trees against a fuzzy

(leafy) background. This example shows that the algorithm

works well with patterns that are more complex than stripes

and geometric patterns.

[0123] In the fourth example of FIG. 7d, the pattern is more
complex and less well de?ned. Nevertheless, the results show
images having a pattern of waves created by a duck or other
bird in water.

[0124] As explained above, the system comprises an image

database and a feature database which stores images and their

associated signatures. The searching described in relation to

FIGS. 7a to 7e can be conducted through the images stored in
these databases. It may be helpful for such databases to be
divided into a library of patterns. For example, a library of
patterns may contain some or all of the following types of

pattern; vertical stripes, horizontal stripes, diagonal stripes,

large dots, small dots, tiger stripes, brickwork, checks and
zig-zags. Clearly other types of categories of patterns could
be used as a sub-section within the library. Each sub-section

has a canonical form for the pattern which is de?ned manually

but as explained in more detail below, the library itself is built

up automatically.
[0125]

Such a library may be helpful when conducting a

iv. The similarity values are then ranked (step S506)

this pattern. E.g., Stripes. (Note that this could be in any

language, making the method language independent). This

label is then assigned to the ranked sub-section (step S508). It
will be appreciated that more than one closest sub-image may
be selected so that the image is tagged with more than one
label. This is dependent on how many sub-sections are within

the library.
[0134] vi. Determine either the dominant color of the image
or of the elements of the pattern (step S510)
[0135] vii. Look up the textual description of the color. E. g.,
Green and Combine the color description and the pattern
description to create an overall description. E.g., Green

stripes (steps S512).

[0136] The ?nal two steps show how the library can be
optionally used to include a color tag. The histogram/signa
ture of the image could be compared with a color histogram to
determine the primary color of the image. FIGS. 8a to 80
show three examples of a canonical form for a sub-section of
a library. In this case, the subsections are red horizontal

stripes, green spots and purple checks.

[0137]

This could be further extended with object detection

to allow the system to label an image as red striped dress or

circular green spotted plate. Thus, the library could contain

sub-sections within the sub-sections for example, horizontal
stripes may be the main category, this may be sub-divided by
color and may be also be sub-divided by object, e.g. horizon

search, for example, when a user does not have an image to

begin the search. In such a case, the user could select the

tal striped dress or green horizontal stripes or even green

pattern they are interested in by clicking on a sample image

showing a canonical form of the pattern.
[0126] FIG. 8d shows how the library is initialized by de?n

[0138] Thus, another method for a user to input a query

image is to type a query such as Striped dress into a text

horizontal striped dress.

ing a representative pattern (canonical form) for each sub

based search engine. The canonical image of stripes may then

be used as the query image. The word dress might be applied

section at step S400. The signature for this sub-section is then

determined as described above (S402). The system deter

as a keyword query alongside the query image or perhaps

mines whether or not there are any further patterns to be

images of dresses from within the images which match the

canonical image for stripes. Alternatively, if a sub-section

included at step S404 and if so repeats the ?rst two steps until

applied against an object recognition database to identify

the library is complete. The library is then output at step S406.

entitled striped dress already exists, the canonical image for

For example, the images shown in FIGS. 8a to 80 may be

this sub-section can be input as the query image to identify

input as the representative pattern for stripes, spots and

checks respectively. It will be appreciated that the library
can be expanded by including or re?ning the subsections by
de?ning new representative patterns and repeating the steps

further images of dresses.

for the method.

[0127] FIG. 8e shows how such a library can be used to

automatically label or tag each image. For example, each

image could be compared with a representative histogram for
each type of pattern. Each image could then be added to the
section of the library with which it has the highest similarity
value or perhaps to more than one section if the similarity
values are above a threshold value.

[0128] FIG. 8e shows that the algorithm to provide the

functionality to tag images using a library could be as follows:

[0129]

i. An image (query image) is received for classi?ca

tion (step S500);

[0130] ii. Compute the histogram or signature of the query
image, as described above (step S502).

[0139]

No doubt many other effective alternatives will

occur to the skilled person. It will be understood that the

invention is not limited to the described embodiments and
encompasses modi?cations apparent to those skilled in the art

lying within the spirit and scope of the claims appended

hereto.
What is claimed is:

1. A non-transitory computer readable medium carrying

computer program code to, when running, implement a
method for processing an image to generate a signature which
is characteristic of a pattern within said image, the method

comprising:
receiving an image;
overlaying a window at multiple locations on said image to

de?ne a plurality of sub-images within said image, with

each sub-image each having a plurality of pixels having

a luminance level;

Jul. 3, 2014

US 2014/0185959 A1

determining a luminance value for each said sub-image,

Wherein said luminance value is derived from said lumi
nance levels of said plurality of pixels;
combining said luminance values for each of said sub

images to form said signature;

Wherein said combining is such that said signature is inde
pendent of the location of each sub-image.
2. The non-transitory computer readable medium accord
ing to claim 1, Wherein determining said luminance value

comprises:
determining an average luminance for each of said plural

ity ofpixels;
determining Whether said average luminance is greater
than a threshold value;

setting each of said plurality of pixels to a ?rst value if said

average luminance is greater than a threshold value;
setting each of said plurality of pixels to a second value if
said average luminance is lower than a threshold value;
and
combining said ?rst values and said second values to gen
erate said luminance value for each said sub-image.
3. The non-transitory computer readable medium accord
ing to claim 2, Wherein said threshold value is the average
luminance of the image.
4. The non-transitory computer readable medium accord
ing to claim 2, Wherein said ?rst value is one and said second
value is zero and said luminance value is a binary number.

5. The non-transitory computer readable medium accord

ing to claim 1, Wherein combining said luminance value
comprises de?ning a histogram having a plurality of slots
With each slot representing a luminance value.

said Window to reduce the number of pixels Within said sub

image to said plurality of pixels.

14. A non-transitory computer readable medium carrying
computer program code to, When running, implement a
method of searching for an image, the method comprising:
receiving a query image;
determining said signature for said query image, Wherein
said signature is generated using the method of any one

of the preceding claims;

retrieving an image from a database together With its cor

responding signature;
comparing said signature for said query image With said
signature for said retrieved image to generate a similar

ity value for said retrieved image;

repeating said retrieving and comparing steps to generate
similarity values for multiple retrieved images;

ranking said similarity values; and

outputting said retrieved images having highest ranks.
15. A non-transitory computer readable medium carrying
computer program code to, When running, implement a
method of creating a database of images, the method com

prising:
de?ning a plurality of sub-sections Within said database
With each sub-section having a corresponding signature

Which is generated by
receiving an image representative of said sub-section;
overlaying a Window at multiple locations on each said

representative image to de?ne a plurality of sub-images

Within each said representative image, With each sub
image each having a plurality of pixels having a lumi
nance level;

6. The non-transitory computer readable medium accord

ing to claim 5, Wherein combining said luminance value
comprises incrementing a slot in said histogram for each
sub-image having a luminance value corresponding to said

determining a luminance value for each said sub-image,

Wherein said luminance value is derived from said lumi
nance levels of said plurality of pixels;
combining said luminance values for each of said sub

slot.

sub-image has 3x3 pixels.

images to form said signature;

Wherein said combining is such that said signature is inde
pendent of the location of each sub-image;
receiving an input image to be input into said database;

8. The non-transitory computer readable medium accord

ing to claim 5, Wherein combining said luminance value

overlaying a Window at multiple locations on said input

7. The non-transitory computer readable medium accord

ing to claim 5, Wherein said histogram has 512 slots and said

comprises converting said histogram to a feature vector to

form said signature.

9. The non-transitory computer readable medium accord
ing to claim 1, Wherein at least some of said plurality of

images partially overlap at least one other image.

10. The non-transitory computer readable medium accord
ing to claim 1, comprising rescaling said received image
before overlaying said Window.
11. The non-transitory computer readable medium accord
ing to claim 1, comprising converting said received image or
said rescaled image to greyscale before determining said
luminance values.

12. The non-transitory computer readable medium accord

ing to claim 1, comprising converting each said sub-image to
greyscale When determining said luminance value for said

sub-image.
13. The non-transitory computer readable medium accord

ing to claim 1, comprising rescaling said sub-image Within

determining a signature for said input image by:

image to de?ne a plurality of input sub-images Within
said input image, With each sub-image each having a
plurality of pixels having a luminance level;
determining a luminance value for each said input sub
image, Wherein said luminance value is derived from
said luminance levels of said plurality of pixels;
combining said luminance values for each of said input
sub-images to form said signature;
Wherein said combining is such that said signature is inde
pendent of the location of each input sub-image;

comparing said signature for said input image With each

signature of said plurality of sub-sections to generate a
similarity value for each of said plurality of sub-sec

tions;
ranking said similarity values; and
allocating said input image to said sub-section having the
highest rank