Visualizing Pattern Constancy

Correlation Coefficients - Residential Scene

Original image/notes Unenhanced Edge Pattern Enhanced Edge Pattern

Almost complete saturation

R=0

R=0.04

Decreasing saturation

R=0.037

R=0.30

.

R=0.22

R=0.50

.

R=0.43

R=0.62

.

R=0.49

R=0.72

Near normal exposure

R=0.48

R=0.79

Best visual representation

R=0.50

R=1

.

R=0.41

R=0.69

Twilight begins

R=0.48 (0.16)

R=1 (0.17)

Darkness deepens

R=0.31

R=0.67

.

R=0.89

R=0.43

.

R=0.37

R=0.35

.

R=0.004

R=0.23

.

R=-0.003

R=0.14

.

R=-0.003

R=0.07

.

-0.003

R=0.03

Nearly all noise

R=-0.001

R=0.02

For a sequence of time series images, we have computed the behavior of the correlation coefficient over a wide
array of varying conditions- near saturation through normal exposures with variable lighting (sun angle mostly,
down to advancing nightfall with increasing sensor noise to its extreme limits). The correlation coefficient is
computed for both un-enhanced edge images and the enhanced edge images to quantify the benefit of enhancement,
per se, and to quantify the degree of pattern constancy that is maintained to " recognize " this specific
scene with a computer under these highly arbitrary imaging conditions.

The chronological order of the images has been rearranged to show the trend from near saturation down to deepening
night. The best enhanced visual representation of the scene is chosen to define the edge image to correlate against
all the other edge images (both un-enhanced and enhanced). The correlation coefficient computed as the covariance
divided by the product of the two separate standard deviations.

       

small logo
Responsible NASA Official: Glenn Woodell

NASA Privacy Statement, Disclaimer, and Accessibility Certification