ADAPTING WEB IMAGES FOR BLIND PEOPLE

A. Fatzilah Misman

Department of Information System, International Islamic University Malaysia, Kuala Lumpur, Malaysia

Peter Blanchfield

School of Computer Science, University of Nottingham, Jubilee Campus, Wollaton Road, Nottingham, U.K.

Keywords: Web accessibility, Image tagging, Describing web image, Semantic web.

Abstract: One way to remedy the gap that evidently exists between the image element on the web and the web user

who is visually blind is by redefining connection between the image and the abundant element of the web

itself i.e. text. Studies on the exploitation are done largely related to the fields like the HCI, the semantic

web, the information retrieval or even a new hybrid approach. However, often many see the problem from

the perspective of the third party. This position paper posits that the problem can also be seen from the

fundamental reasons for an image being on a web page without neglecting the connection that develops

from the web user’s perspective. Effective and appropriate image tagging may consider this view.

1 INTRODUCTION

Adapting web images to make them useful to blind

users of the Internet presents a major technological

challenge. In practice web images are commonly

ignored by blind users due to the considerable

difficulty in obtaining meaningful output. Evidence

for this has been found from a preliminary study

with users of a local blind centre. This is also

supported by evidences from related studies (Petrie,

O’Neill & Colwell, 2002, Petrie & Kheir, 2007).

However, “Meaning can be as important as

usability in the design of technology (Shinohara &

Tenenberg, 2009)”. Thus theoretically the ‘meaning’

of an image is bounded by values of information an

image can offer. King, Evans, and Blenkhorn (2004)

indicated that there are four common technological

solutions for blind users to access the Internet

including screen reader technology working

alongside a human assistant to interpret images for

them. When using a screen reader most users will

have to listen to the whole of an image tag and try to

interpret the meaning of the image from the content

of the tag. It is seldom possible for them to gain the

same level of meaning from an image by this means

as would be possible for a sighted user looking at the

image. If an “alt” tag has been used the system may

choose to read this alone. However, the purpose of

an image is often not well defined by the content of

these “alt” tags (Bigham, Kaminsky, Ladner &

Danielsson, 2006, Petrie, Harrison & Dev, 2005).

The ability to re-tag images with appropriate

descriptions is the main goal of this paper. There

have been various other attempts to retag images but

this paper proposes a novel approach to this

retagging which deliberately emphasises the actual

use which blind users will make of the images rather

than merely trying to retag on the basis of inferred

knowledge from the tags. For example King et al.

(2004) demonstrated this by developing a dedicated

web browser for blind users. They suggested that

images can be ignored entirely except when they

contain valuable information in the “alt” tags or

when containing a hypertext link which gives

meaningful information with regard to a hypertext

destination. However, in the current research an

ethnographic study has found that blind users will

want to retain images if they have direct relevance to

the content. Thus, we believe, if the image tag is

exploited effectively there are circumstances where

they can represent better information when read

using a common screen reader. Once the relevance

of an image has been established the blind user will

then often gain the help of a sighted user to explain

the purpose of the image.

430

Fatzilah Misman A. and Blanchﬁeld P..

ADAPTING WEB IMAGES FOR BLIND PEOPLE.

DOI: 10.5220/0003402004300437

In Proceedings of the 7th International Conference on Web Information Systems and Technologies (WEBIST-2011), pages 430-437

ISBN: 978-989-8425-51-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

There are a vast number of images on the World

Wide Web today. However, the standard guidelines

for providing appropriate descriptions for them are

not enforceable (Caldwell, Cooper, Reid, &

Vanderheiden, 2008). Despite numerous studies on

visual processing of images to produce annotation or

explanation, there is as yet no software that can

determine image content in a widely useful way

(Russell, Torralba, Murphy, & Freeman, 2005, Li &

Wang, 2008). In addition to image analysis

approaches a number of other studies have been

undertaken that make use of data mining principles

specifically to produce good image tagging for

images (Bigham et al., 2006, Bigham, 2007) The

position of the current study is to establish a

definition of a “meaningful image” for blind people

and thus enable valuable retagging of images to take

place. It embodies expectations expressed by blind

people and exploits the role an image plays on the

page it is in.

Current techniques for generating image

descriptions based on text content are insufficient in

many ways. It is generally assumed that elements in

a web content (particularly body text and images

that appear within the same page or within the same

virtual boundary such as a title or a head line) are

associated to one another. This presumption has

however always been challenged – see for example

Carson and Ogle (1996). They asserted that even at

that stage images might not be related to the content

of a page. For example an advertisement image may

have no direct relationship at all to the content of a

page. This concern may be more significant for

issues dealing with information retrieval but may not

be as important when dealing with providing access

to information on images on a page for blind users.

2 RELATED WORK

Previous studies on generating image labels or

descriptions have made use of text descriptions via

the html’s “longdesc” or “alt” syntaxes for image

descriptions. Despite the significant amount of work

that has been done the only reliable method

available for providing precise description to this

day is manual labelling (Ahn & Dabbish, 2004). The

reasons are that this method is independent,

unrestricted, self manageable and original especially

when it is performed during the development phase

when the author initially is adding image elements to

the web site. The problem however is that this

approach may become tedious and thus extremely

costly, particularly when the web page has already

been published.

As an alternative solution, studies on image

description have looked into automatic approaches

to labelling. Ahn and Dabbish (2004) developed a

word-image-matching online game that exploited

players contributions to generate descriptions for

web images. The game involves pairs of players

who are randomly partnered to guess what each

other have keyed in for the same image. Every time

both players describe an image similarly they can

then move on to the next image and the word or

phrase used is chosen to describe the given image.

This method adapted the concept implemented in the

“Open Mind Initiative” project (Stork, 1999, Stork

& Lam, 2000). The Open Mind Initiative applied the

intelligent agent concept where the machine is

trained about the attributes of various sets of image

groups derived from a training database which was

also contributed to by public users. Outsourcing data

from open contribution conveniently creates

sufficiently representative training databases (Datta,

Li & Wang, 2005). However, use of outsourced

labelling fails to address fundamental issues of web

image dynamism. For example, images from a web

source can be interpreted differently by different

people. Also, for example with news web sites,

images are added to sites frequently and in large

numbers. In addition a huge number of new pages

are added to the Web daily.

However, automatic techniques which have

built-in mechanisms to generate image descriptions

are also under development. For example the ALIPR

(Automatic Linguistic Indexing of Pictures – Real

Time) used a large database consisting of web pages

as a training resource to develop learning algorithm

of image signatures (Li & Wang, 2008). The

learned patterns cluster images discretely and

categorise from binary text to image pixels. The

clustered patterns are then used to produce real time

image descriptions. However, dependence on a large

database system would introduce the issue of

reliability of the approach when running on sizable

data.

Bigham et al. (2006) use a web domain that runs

a combination of three methods, enhanced web

context labelling, Optical Character Recognition

(OCR) and human labelling, to perform the labelling

task. The use of three independent systems has the

effect of increasing processing load. The web

context labelling utilises summaries of the page, the

title or header to imply the image content. This

system also uses the content of linked pages for the

same purpose.

ADAPTING WEB IMAGES FOR BLIND PEOPLE

431

The Latent Semantic Analysis (LSA) was

originally used to reduce the semantic

dimensionality of information retrieval problems

such as the use of the synonym concept to elaborate

an object query (Deewester et al., 1990). The theory

of the LSA has been widely used and has evolved

into varieties such as the Explicit Semantic Analysis

(ESA) by Gabrilovich and Markovitch (2007). The

ESA particularly computes semantic relatedness

among words of human natural languages by

comparing the corresponding vectors using

conventional metrics (e.g., cosine). The ESA

experiment on Wikipedia®-base knowledge data

demonstrated substantial improvement based on the

yield assessment. It also demonstrated that, limited

words such as the information tagged to an image

could yield more relevant words which can thus be

used to describe the image. The problem however,

is that this approach is not suitable when small

databases have to be used and so new approaches

must be found.

3 PROPOSED APPROACH

The proposed approach has been influenced by the

results of an ethnographic study as mentioned in the

introduction. This study looked at the use being

made of images by blind people when using the

web. The study took place in a local centre for the

blind, which the users attend to gain help in

accessing the web and its content. It was clear from

this work that general use of the internet by the

blind was for similar purposes to that of sighted

people (Shinohara & Tenenberg, 2007). The centre

provided screen readers and the addition of sighted

assistants for helping the blind users. Observations

were made through the use of general conversations

and later from direct questioning about the value or

otherwise of image content. By understanding the

real scenario and the expectation of blind people

from web images they come across, the question

arose as to whether or not images could be adapted

to become meaningful for their benefit using the

existing independent screen reader software such as

JAWS® and SuperNova®. The responses were

positive provided web images had readable

descriptions that at least met one of the following

criteria:

 Conciseness – that is the tag is concise and

gives an accurate reflection of its relationship

to the page content.

 Appeal – the image tag invites the reader to

continue to read the page content.

 Readability - any understandable description

carries value.

Specifically these ideas were derived from the

experiences in the ethnographic study. During this

process two approaches were adopted. In the first

the sighted helper became directly involved in the

browsing process, responding to requests from the

blind person while they listened to the screen reader.

When images were viewed their relevance would be

assessed by the helper. If the tagging was too long or

inappropriate the helper would provide their own

interpretation. In the second the browsing is

directed by the blind user. The helper would correct

them if they went wrong in the process. The blind

user would then ask help when they noticed a word

that seemed relevant in the screen reader output.

They would then request assistance from the helper.

Often the blind user would be misinterpreting

keywords in an image tag. However, they would

mostly skip most of the image elements. A similar

observation with a single user was made by

Shinohara and Tenenberg (2007).

This paper proposes an approach that combines

i) Image purpose ranks developed from the work of

Paek and Smith (1998) on image purposes ii) an

interpretation engine which filters and replaces the

HTML within the image tag with the relevant

semantic direct interpretation and iii) a Proximity

Weighting Model (PWM) that weighs replacement

tags by measurement of relevant phrases or

sentences within close proximity (that is, page

proximity). All of these approaches have been used

before but in this case the result which is delivered

to the blind user is interpreted with respect to the

analysis of purpose results derived from the work

with blind users. So instead of replacing an image

with text, for example, its value to a blind user is

first assessed. If the image is content related (within

the definitions given in Table 1 below) then it will

be retained but retagged, clearly and concisely

indicating it is content. If it is purely navigational

the tag should be adapted to reflect its navigational

purpose but the image information (contained for

example in the ‘<img>’ tag) might be removed.

This paper proposes that text within a web page

retains a certain level of association to other

elements within the page in various ways. Thus,

adapting images for blind user means to generate a

description tag which fulfils at least one if not all the

three criteria stated above.

3.1 Image Purpose Rank

According to (Sutton & Staw, 1995) the ‘core’

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

432

question for an explanation is the question ‘why’. It

equips an occurrence with a sense of reason. Paek

and Smith (1998) suggested a technique to improve

the cataloguing and indexing of web images based

upon image contents. They identified image

purposes associated to image content. They

suggested that every embedded web image could be

categorised as content, navigation, decoration, logo,

advertisement, information and content as described

in Table 1.

Table 1: Adapted from Paek and Smith (1998) purposes of

web image.

Purpose Definitions

Advertisement Image that contributes to the act of

informing, notifying or promoting

what it represents which may or may

not be related to the page's content.

(e.g. corporate logo, individual brand

on an anonymous website)

Decoration Image that is meant for decorating the

page. (e.g. buttons, balls, rules,

masthead, background)

Information Image that signifies the message

which it represents. (e.g warning

signs, under construction, what's new,

graph figures)

Logo Graphic representation or symbol of a

company name, trademark,

abbreviation, etc., often uniquely

designed for ready recognition. (e.g.

IBM, corporate logo)

Navigation Image that represents the act or

process of navigating within or

without the page. (e.g. hyperlinked

image, arrows, home, image map)

Content Image that is associated with a body

of text of a source page. (e.g, Honda

advert on Honda homepage, image

that is transcribed in the page content)

In order to verify that the current authors’

interpretation of this claim was correct an online

survey with a group of expert users whose

professional experiences are related to Information

Technology was made. Fifteen respondents took part

where each assessed a randomly selected and

reconstructed web page from the original source. Six

images from each page were indexed with the

respective question number. Each question states

seven options to range from the six image purposes

and one ‘uncertain’ option. The results of this

survey were convincing that this interpretation was

correct and so could be used for the classification

process. Sixty three percent of the responses exactly

matched those of the authors. Of the other responses

there was a range of disagreement between the

correspondents but the majority were generally more

in agreement with the authors.

Next, image purpose profiling was developed to

define the image classification categories. The

classifications were developed based on a study of

patterns derived from a total of 1707 images from

175 web pages chosen using Yahoo!® API. The

websites were from fairly mixed domains consisting

of organization, government, commerce, social

networking and educational websites. The

classification defines regular patterns of occurrences

within image tags identified by keywords, HTML

syntax and frequency of occurrence of each

keyword. The identified set of keyword patterns

were then considered as a set of definition criteria

representing the respective image’s purpose. Every

keyword is given a weight that indicates its

contribution to an image’s purpose’s on a web site.

For example, the keyword ‘ads’ has a mean score of

74% matching to the advertising category and the

keyword ‘icon’ has a mean score of 78% indicating

it as a decorative image.

Therefore, the purpose ranking role is to generate

semantic expression which indicates ranked

purposes definition based on the ranking

perspectives algorithm. Figure 1 shows an example

result of semantic expression of purpose using this

ranking.

Figure 1: Example of semantically expressed purpose

rank- “This image is quite likely navigational

and a

content’s

less likely an advertising or a decoration or the

least a logo

or an information”.

The algorithm basically works by calculating the

accumulated mean weight of any identified pattern

found within the image tag and dividing it by the

number of pattern occurrences (see Table 2).

3.2 Semantic Interpretation of an

Image Tag

While HTML image taglines carried by image

function tags ‘<img/>’, ‘<map>’ and ‘<area/>’

ADAPTING WEB IMAGES FOR BLIND PEOPLE

433

Table 2: Semantic rank on purpose point score.

Semantic

rank

likely

Quite

likely

Less

likely

The

least

Mean

score

rank

75 – 99

(%)

50 – 74

(%)

25 – 49

(%)

1 – 24

(%)

Mean

scores of

Figure 1.

Navigat

= 0.6

Content’s

= 0.56

Advert

= 0.4

Decor

= 0.3

Logo

= 0.2

Info

= 0.2

have the intended interpretation for web developer’s

purposes their interpretation by web users would be

far less obvious. However, blind users using a

screen reader will encounter the whole of the image

tag in place of the image. The aim of the current

work is to replace these tags where possible with

tags that more concisely represent the purpose of the

tag. Thus as said earlier if a tag is merely

navigational it can be replaced by an appropriately

tagged link. However if it is content related it must

get a suitably concise, appealing and readable tag.

Table 3: Example of image tag attributes.

Required Attributes

Attribute Value Description

alt Text Specifies an alternate text

for an image

src URL Specifies the URL source

of an image

Optional Attributes

height Pixels% Specifies the height of an

image

longdesc URL Specifies the URL to a

document that contains a

long description of an

image

usemap #mapname Specifies an image as a

client-side image-map

width Pixels % Specifies the width of an

image

The original image tag attributes provide

information from which the interpretation of the

image’s purposes can be inferred and their

acceptable values can be determined. The image tag

attributes are structured and designed for the

developer’s convenience. This feature is particularly

interesting for the case of blind users whereby image

tags can possibly be used for semantic interpretation

answering three questions related to the image. They

are;

 What does the image do?

 Where does it come?

 What is the image possibly about?

Possible viable answers are contained within

‘href=’, ‘src=’ and ‘alt=’ as presented in Table 3.

Table 4: Examples of image tag interpreted attribute.

Img tag 1 <img src="penguin-pictures-2.jpg"

alt="penguin pictures" width="775"

height="271">

Interpret 1 image from penguin pictures jpg file

.noted as penguin pictures

Img tag 2 <a href="http://www.penguin-

pictures.net/adelie_penguin_pictures.html

"> <img src="adeliepenguinpictures.gif"

alt="adelie penguin pictures"

width="165" height="30" target=

“blank”></a>

Interpret 2 navigates to web site penguin pictures

network adelie web page .image from

adeliepenguinpictures gif file .noted as

adelie penguin pictures.on new window

3.3 Proximity Weighting Models

(PWM)

PWM is often applied in the first stage of the Latent

Semantic Analysis (LSA), as part of the local weight

function (LWF) (Nakov, Popova & Mateev, 2001).

It enhances document retrieval effectiveness by

examining documents in terms of local level

statistics (Macdonald & Ounis, 2010). Modern

information retrieval systems generally take more

than a single-term weighting model to rank

documents. The PWM has contributed a significant

factor in addition to LSA assessment for context

relatedness within a document (Croft, Metzler &

Strohman, 2009). It is based on the theory which

states that proximity or distance between words

retains certain weight of connection. For example, a

repetition of the same noun though positioned in

different lines or in different paragraphs of a

document, suggests that the word is important to

those lines or to the paragraphs.

The objective of PWM application in this case is

to measure relatedness of page content text by

measuring their weight distance from the image. The

weighting scheme and computation is adopted from

the model presented by (Cascia, Sethi & Sclaroff,

1998, Sclaroff, Cascia & Sethi, 1999). This was

selected because of the implication of the likelihood

of useful information occurrences that may be

connected to an image tag. Besides, this method is

sensitive to the distance of an individual pair

depending on the way aggregation occurs locally

(Tao & Zhai, 2007).

To obtain result quality, preprocessing of the

page document is performed i.e. removing stop

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

434

words and stemming. Each image’s associated

HTML tag is parsed and a word frequency

histogram is computed such as in (Table 5).

Table 5: Terms from image tag. Cell entries are the

number of times that words (column) appeared in the

content’s HTML tags (row).

Phrases containing associated words are not

always similar in length and structure. Words

appearing with specific HTML tags are considered

special thus being assigned with higher weight than

all other words - see Table 6.

Table 6: Cascia et al.’s (1998) word weight scheme on

HTML tags.

These weight values are heuristically derived

based on selective weighting of words appearing

between various HTML tags that measure their

estimated useful information implied to the text. The

importance given to a word is computed from its

frequency, location relative to tags as shown in table

6 and relative to its proximity to the image (as

described below). For example, if the word ‘adelie’

appears within ‘<bold>’ tags in the HTML once,

then it is initially weighted at 3.0. If instead it

appears in text without HTML tags it is initially

weighted as 1.0. Multiple occurrences also carry

multiplier to the single value.

For words appearing before and after a particular

image tag, the proximity weighting value is

computed from equation (1) where pos is the

position of the word with respect to the image tag

and dist is the maximum number of words

considered in applying such weighting. In this

implementation, dist is 10 for words appearing

before the image and 20 for words appearing after

the image. The constant

= 5.0 is assigned so that

the nearest text to the image tag is slightly less than

and equal to the words appearing in the alt field of

that image and the ‘title’ of the page respectively.

2.0 /

os dist

−•

⋅

(1)

This weight distance proximity function is based

on the assumption that words close to an image have

connection to the image. It is also assumed that the

degree of the connection decreases exponentially

with the distance of the recurring word from the

original word in the image although this may not

always be the case. This assumption is founded in

reference to the analytical survey results made by

Cascia et al. (1998) themselves and by Sclaroff et al.

(1999).

Furthermore, for this paper the adopted

proximity weighting scheme is not used to represent

context association for the LSA. The scheme is

rather to generate viable context from the HTML

text association within the tags or a sentence as a

whole. The objective is to justify the selection of the

sentences or phrases from the HTML text hence

used to add to the appeal factor in the automatic

image tagging.

4 CONCLUSIONS

The results produced thus far look potentially

promising when run on 20 sampled websites. The

application produced three layered descriptions for

every image successfully although semantically the

value of each varied. However, this preliminary

indication has not been proven with the real users

which should provide data that help measure the

methodological significance. Furthermore, the

consistency of the results seems to be affected by the

weight of text proportion as well as the design

quality of the page based on the Web Accessibility

Initiative (WAI) standard benchmark.

An experiment to verify this approach is being

undertaken in two stages. 1) Control experiments on

normally sighted users, which will attempt to

evaluate the generated description based on

feedback from the sighted users. This will be used as

the benchmark measure for the result of the second

stage. 2) Acceptance experiment on the target group.

This experiment will attempt to evaluate the real

users experience by comparing between the pages

ADAPTING WEB IMAGES FOR BLIND PEOPLE

435

with and without the adaptation applied. This will

provide information which may prove the

significance of the hypothesis of this study.

The position of this paper is that images on web

sites need to perform their intended purpose for

blind users as much as for sighted users. It is

therefore inappropriate to simply remove image tags

as suggested by some (Shinohara & Tenenberg,

2009). It is also more important to present the

purpose of images concisely and in an appealing

way to the blind user. In some cases this may result

in the replacement of the image with a link – when

the image is merely being used as navigational cue.

However, when the image is part of the page’s

information it is necessary to retain the image and

introduce an alt tag that explains clearly and

concisely the functional relationship of the image to

the text. So for example for an image of a car in a

used car advertisement, if the image is of the actual

car it will be retained and the alt tag replaced with a

phrase such as “An image of the car advertised.”

This then allows the blind user to know the purpose

of the image and obtain help from a sighted person

in determining the value of the content. However, if

for example the image is a generic one it might be

removed.

REFERENCES

Bigham, J., Kaminsky R., Ladner, R., Danielsson, O., &

Wempton, G. (2006) Webinsight: Making web images

accessible. In Proceedings of 8th International ACM

SIGACCESS Conference on Computers and

Accessibility, 181-188.

Bigham, J. (2007) Increasing Web Accessibility by

Automatically Judging Alternative Text Quality. In

Proceedings of IUI 07 Conference on Intelligent User

Interfaces, 162-169.

Caldwell, B., Cooper, M., Reid, L., & Vanderheiden, G.

(2008) Web Content Accessibility Guidelines 2.0,

Retrieved September 23 , 2009 from

http://www.w3.org/TR/WCAG20/

Carson, C. & Ogle, V. E (1996) Storage and Retrieval of

Feature Data for a Very Large Online Image

Collection. In IEEE Computer Society Bulletin of the

Technical Committee on Data Engineering, 19(4).

Cascia, L, Sethi, S. & Sclaroff, S. (1998) Combining

textual and visual cues for content-based image

retrieval on the world wide web. In IEEE Workshop on

Content-Based Access of Image and Video Libraries.

Croft, B., Metzler, D. & Strohman, T. (2009) Search

Engines: Information Retrieval in Practice. Addison-

Wesley.

Datta, R., Li, J. & Wang, J. Z. (2005) Content-based

image retrieval: approaches and trends of the new age.

In Proceedings of the 7th ACM SIGMM International

Workshop on Multimedia Information Retrieval, 253–

262.

Deewester, S. Dumais, S., et al. (1990) Indexing by latent

semantic analysis, Journal of Americ. Soc. Info. Sci.,

vol 41, 391-407.

Gabrilovich, E. & Markovitch, S. (2007).

Computingsemantic relatedness using Wikipedia-

based explicit semantic analysis. In proceedings of

IJCAI, 1606-1611.

King, A., Evans, G., & Blenkhorn, P., (2004) WebbIE: a

web browser for visually impaired people. In S.

Keates, P. J. Clarkson, P. Langdon, & P. Robinson

(Eds)., In Proceedings of the 2nd Cambridge

Workshop on Universal Access and Assistive

Technology (pp. 35 – 44), Springer-Verlag, London,.

Li, J. & Wang, J. Z. (2008). Real-time computerized

annotation of pictures. IEEE Trans. Pattern Anal.

Mach. Intell. 30(6), 985–1002.

Macdonald, C. & Ounis, I. (2010) Global Statistics in

Proximity Weighting Models. In Proceedings of Web

N-gram Workshop at SIGIR 2010, Geneva.

Nakov, P., Popova, A. & Mateev, P. (2001) Weight

functions impact on LSA performance, In Proceedings

of the Euro Conference Recent Advances in Natural

Language Processing (RANLP), 187–193.

Paek, S. & Smith, J.R.(1998) Detecting Image Purpose. In

World-Wide Web Documents. In Proceedings of

IS&T/SPIE Symposium on Electronic Imaging:

Science and Technology – Document Recognition. San

Jose: CA.

Petrie, H., Harrison, C. & Dev. S. (2005) Describing

images on the web: a survey of current practice and

prospects for the future. In Proceedings of Computer

Human Interaction International (CHI).

Petrie, H., O’Neill, A-M. and Colwell, C. (2002).

Computer access by visually impaired people. In

A.Kent and J.G. Williams (Eds.), Encyclopedia of

Microcomputers, vol 28. New York: Marcel Dekker.

Petrie, H. and Kheir, O. (2007) The relationship between

accessibility and usability of websites. In Proceedings

of Computer Human Interaction International (CHI)

Russell, B., Torralba, A., Murphy, K., & Freeman, W.,

(2005) LabelMe: a database and web-based tool for

image annotation (Technical Report). Cambridge,

MA: MIT Press.

Sclaroff, S., Cascia, L. & Sethi, S. (1999) Using Textual

and Visual Cues for Content-Based Image Retrieval

from the World Wide Web, Image Understanding,

75(2), 86-98.

Shinohara, K., & Tenenberg, J. (2007). Observing sara: a

case study of a blind person’s interactions with

technology. In Proceedings of the 9th international

ACM SIGACCESS conference on Computers and

accessibility. New York: ACM Press. 171–178

Shinohara, K. & Tenenberg, J., (2009). A blind person’s

interactions with technology, Communications of the

ACM, 58-66.

Stork, D. G. (1999) The Open Mind Initiative. IEEE

Intelligent Systems & Their Applications, 14(3), 19-20.

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

436

Stork, D. G. & Lam C. P. (2000) Open Mind Animals:

Ensuring the quality of data openly contributed over

the World Wide Web. AAAI Workshop on Learning

with Imbalanced Data Sets, pp 4-9.

Sutton, R. I & Staw, B. M. (1995) What theory is not

Administrative Science Quaterly, vol 40, 371-384.

Tao, T. & Zhai, C. (2007) An exploration of proximity

measures information retrieval. In Proceedings of

SIGIR’ 2007, 295-302.

Von Ahn, L. & Dabbish, L. (2004). Labeling images with

a computer game. In Proceedings of the ACM SIGCHI

Conference on Human Factors in Computing Systems

(CHI).

ADAPTING WEB IMAGES FOR BLIND PEOPLE

437