Saliency-based color accessibility

Saliency-based color accessibility

Satohiro Tajima*, and Kazuteru Komine

*[email protected]

1-1.3% 0.02% 1.3% 0.02%

1-1.2% 0.01% 5.0% 0.35%

0.001% 0.03% 0.01% 0.01%

Divergence of color perception

Protanope ∙ ● ●

Deuteranope ● ∙ ●

Tritanope ● ● ∙

L M S

Male Female Male Female

Dichromacy Anomaloustrichromacy

Color deficiency ~280 million (8% of male)

Common ● ● ●

Saliency-based color accessibility

What you see ≠ What another one sees.Original image

Deuteranopia (simulation)

Saliency-based color accessibility

Conventional simulation methodOriginal Protanope

Tritanope Deuteranope

(Viénot et al., Nature, 1995)Saliency-based color accessibility

Conventional simulation method

• No objective criteria

• Simulated color distance Perceptual difference??

(Brettel et al., JOSA A, 1997)

Problems:Original

Tritanope

(Viénot et al., Nature, 1995)Saliency-based color accessibility

“Saliency-based accessibility”

• Objective criteria

• Information theoretic distanceWe propose:

Saliency-based color accessibility

Saliency: model of visual attention

(Itti & Koch, Nat. Rev. Neurosci., 2001)Saliency-based color accessibility

Our method: (1) Extended visual saliency model

Observer model

Red—Green

Blue—Yellow

Luminance

Saliency-based color accessibility

Our method: (2) Loss of salienceDeuteranopeControl

Difference:Lost information

Saliency-based color accessibility

Information theoretic measure of loss

L(r) ≡ {lnP(r|Si) – lnP(r|Sj)}

• Weighted log-difference:

DKL[P(r|Si)||P(r|Sj)] = Σr P(r|Si) L(r)• KLD (Kullback–Leibler divergence):

‘Local divergence’

…Global loss

…Local loss

r: pixel locus, S: subject modelSaliency-based color accessibility

L

Model behavior

Saliency-based color accessibility

L

Model behavior

Saliency-based color accessibility

Experiment

1) Original vs. Dichromat simulation 2) Original vs. Saliency-based improve

• Participants: 5 color-deficient, 18 control observers

• Task: Comparison of visual clarity, 2AFC– W/ and w/o spatial context:

Time

Saliency-based color accessibility

Result

W/ context W/O context

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1ControlCD

P(s

elec

ted

as “

bett

er”)

Dichrom. simul.

Chance

**

**

**

**

* p < .05

** p < .001

** **

Saliency improveW/ context W/O context

*

**

n.s.

** *

Improved pictures were favored more often than simulated onesfor every conditions by both groups (p < .0001).

Saliency-based color accessibility

Result: model vs. data

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4 r = 0.46, p = 0.0051y = 0.17x + 0.022

Lost information

P(s

elec

ted

as “

bett

er”)

[con

trol

,dic

hrom

.sim

ul.,

w/o

con

text

]

Saliency-based color accessibility

Applications

Fixed

Original

Saliency-based color accessibility

Message

New method to analyzecolor accessibility was

“Saliency-based accessibility”

Extensions:• Color/non-color deficit by aging (e.g., cataract)• Color gamut reduction• Sound accessibility auditory saliency

proposed,experimentally evaluated, andapplied to TV program.

What you see ≠ What another one sees What you attend = What another one attends

Saliency-based color accessibility

Reference:Satohiro Tajima, and Kazuteru Komine.Saliency-based color accessibility.IEEE Transactions on Image Processing, 24(3), 1115-1126, (2015).

Saliency-based color accessibility