26
WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

  • View
    218

  • Download
    0

Embed Size (px)

Citation preview

Page 1: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

WebInSight:Making Web Images Accessible

Jeff Bigham

Richard LadnerRyan KaminskyGordon HemptonOscar Danielsson

Page 2: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Some statistics

10 million blind people in the U.S. 55,000 blind children 5 million blind people over 65

Computer and Internet Use 1 million use computers < 200,000 have access to the Internet < 100,000 use a computer regularly 32% of legally blind adults employed

Source: American Foundation for the Blind Blindness Statistics http://www.afb.org/Section.asp?SectionID=15

Page 3: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Browsing the web while blind

Blind users use screen readers Alternative text is substituted for images When no alternative text provided

nothing filename (060315_banner_253x100.gif) link address

W3C accessibility standards “Provide a text equivalent for every non-text element” For images with purely visual purpose, a text equivalent is

an empty string

Page 4: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson
Page 5: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson
Page 6: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Outline

Web Studies

WebInSight System

Where Labels Come From

Evaluation

Future Work

Page 7: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Web Studies Images can be significant or insignificant Significant images need alternative text

alt, title, and longdesc HTML attributes Insignificant images need empty

alternative text(spacers, lines, wacky backgrounds, etc.)

Significance from size, color and function

Page 8: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Web Studies: Groups

CSE Traffic 1 week. 11,989,898 images. 40.8% significant

63.2% assigned alternative text

Popular/Important Websites 500 High-Traffic International Sites 100 Top International Universities 158 Computer Science Departments 137 Federal Agencies 50 States plus District of Columbia

Page 9: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Study Results

Group Significant Insignificant > 90% Number

High-traffic 39.6 27.4 21.8 32913

Computer Science

52.5 41.6 27.0 4233

Universities 61.5 70.2 51.5 3910

U.S. Federal Agencies

74.8 66.6 55.9 5902

U.S. States 82.5 77.1 52.9 2707

Page 10: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Result Graphs

U.S. Federal Agencies

0

20

40

60

80

100

0 20 40 60 80 100 120 140

High-Traffic Websites

0

20

40

60

80

100

0 50 100 150 200 250 300 350 400 450 500

Universities

0

20

40

60

80

100

0 20 40 60 80 100

Comp. Science Departments

0

20

40

60

80

100

0 20 40 60 80 100 120 140 160

Page 11: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Outline

Web Studies

WebInSight System

Where Labels Come From

Evaluation

Future Work

Page 12: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

WebInSight

Add alternative text as a user browses

Coordinate multiple labeling sources

Avoid harming the user experience

Maintain security and privacy

Page 13: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Database

WebInSight Architecture

The InternetThe Internet

Context Labeling

OCR Labeling

Human Labeling

TransformationProxy

GET http://www.cs.washington.eduGET http://www.cs.washington.edu/

GET http://www.cs.washington.edu

Login: _______

Pass: _______

Login: _login___

Pass: _pass___

Page 14: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson
Page 15: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

WebInSight as a Proxy

Transformation proxy Inserts alternative text into webpages Inserts AJAX hooks to allow later changes

Advantages Centralized control Simple setup and administration

Disadvantages Potentially a bottleneck Less control over user interface Secure connections don’t benefit or are less secure

Page 16: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Outline

Web Studies

WebInSight System

Where Labels Come From

Evaluation

Future Work

Page 17: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Providing Labels: Context Labeling Many important images are links

Linked page often describes imageFunction much better than nothing

Page 18: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Providing Labels: OCR Labeling

Original image not recognized(No Text)

Find major colors

Highlight major colors & try again

Page 19: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Providing Labels: OCR Labeling 2

Color Image OCR Text(No Text)

(No Text)

(PIC)t

(No Text)

,, ., ,, ,. ,., , ,,, .,,,n(PIC)(PIC)

Register now!

Page 20: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Providing Labels: OCR Evaluation

Tested 100 images containing text The OCR correctly labeled 52 Our processing correctly labeled 65

Page 21: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Providing Labels: Human Labeling

Humans are best labelers

Luis von Ahn’s games get people to do it

WebInSight sends images to such services

Page 22: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Outline

Web Studies

WebInSight System

Where Labels Come From

Evaluation

Future Work

Page 23: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Evaluation

Experiment Run WebInSight on pages from Web Studies 43.2% of unlabelled sig. images labelled Of these, 94.1% were correct

Page 24: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Evaluation:UCLA

Page 25: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Future Work

User Studies Does it help? What do user’s want out of alt text? When should WebInSight provide it?

Refactoring alt text Present alt text in the best way possible for users

Tool for Webmasters People will always be better but they need help

Page 26: WebInSight: Making Web Images Accessible Jeff Bigham Richard Ladner Ryan Kaminsky Gordon Hempton Oscar Danielsson

Demo