Making Web Images Accessible

1
Making Web Images Accessible
Jeffrey P. Bigham Richard Ladner, Ryan Kaminsky,
Gordon Hempton, Oscar Danielsson University of
Washington Computer Science Engineering
2
Browsing while blind

• Screen readers
• Images cannot be read
• W3C accessibility standards
• Provide a text equivalent for every non-text
  element
• What if no alternative text?
• Nothing
• Filename (060315_banner_253x100.gif)
• Link address (http//www.cs.washington.edu)

3
(No Transcript)
4
nav_svcs.gif
5
Outline

• Web Studies
• Providing Labels
• WebInSight System
• Evaluation
• Future Work

6
Web Studies All Images !

• Significant images need alternative text
• Informative
• alt, title, and longdesc HTML attributes
• Insignificant images need empty alt text
• Automatic Determination?

titlesales graph longdescsales_descrip.txt
height"1"
alt

• More than one color AND both dimensions 10
  pixels
• An associated action (clickable, etc.)

7
Web Studies

• Previous studies
• img tags with defined alt attribute
• 27.91, 47.72, and 49.42
• Significant images have a defined alt attribute?
• 76.93
• Gaps
• Some Ignore Image Significance
• Some Ignore Image Importance

1 T. C. Craven. Some features of alt text
associated with images in web pages.
(Information Research, Volume 11, 2006). 2
Luis von Ahn et al. Improving accessibility of
the web with a computer game. (CHI 2006) 3
Helen Petrie et al. Describing images on the
web a survey of current practice and prospects
for the future. (HCII 2005)
8
Web Studies

• University of Washington CSE Department Traffic

• 1 week
• 11,989,898 images.
• 40.8 significant
• 63.2 alt text

Significant images with alternative text.
Significant images without alternative text.
9
Study Results
Percentage of significant images provided
alternative text, pages with over 90 of
significant images provided alternative text,
number of web sites in group, and number of
images examined.
10
Outline

• Web Studies
• Providing Labels
• WebInSight System
• Evaluation
• Future Work

11
Providing Labels Context Labeling

• Many important images are links
• Linked page often describes image
• What happens if you click

altPeople of UW
People of UW h1People
12
Providing Labels OCR Labeling
(Optical Character Recognition)
Improves recognition 25 relative to base OCR!
4 Jain et al. Automatic text location in
images and video frames. (ICPR 1998)
13
Providing Labels Human Labeling
5
6

• Humans are best
• Recent games compel accurate labeling
• WebInSight database has over 10,000 images
• Could do this on demand

5 Ahn et al. Labeling images with a computer
game. (CHI 2004) 6 Ahn et al. Improving
the accessibility of the web with a computer
game. (CHI 2006)
14
Outline

• Web Studies
• Providing Labels
• WebInSight System
• Evaluation
• Future Work

15
WebInSight System

• Tasks
• Coordinate multiple labeling sources
• Insert alternative text into web pages
• Add code to insert alternative text later
• Features
• Browsing speed preserved
• Alternative text available when formulated
• Immediate availability next time

16
The Internet
Blind User
17
Outline

• Web Studies
• Providing Labels
• WebInSight System
• Evaluation
• Future Work

18
Evaluation

• Measuring System Performance
• WebInSight tested on web pages from web studies
• Used Context and OCR Labelers
• Labeled 43.2 of unlabeled, significant images
• Sampled 2500 for manual evaluation
• 94.1 were correct
• Proper Precision/Recall Trade-off

19
Evaluation Demo
20
Conclusion

• Lack of alternative text is pervasive
• WebInSight calculates alternative text
• WebInSight inserts alternative text
• High precision and moderate recall

21
Future Work
Users
Content Producers

• User Studies
• What do users want?
• How can we provide it?
• Maintain experience.

• User Studies
• Designer motivation.
• Tools for Web Design
• People can always be better
• Adapt user techniques

Common Themes

• Technology
• Improved labeling
• Bring closer to user
• Move beyond images

• More challenges
• Content Structure
• Dynamic Content
• Web applications

22
WebInSight http//webinsight.cs.washington.edu
Thanks to Luis von Ahn, Scott Rose, Steve
Gribble and NSF.