7
Producing a color target acquisition metric Assaf Asbag, Racheli Hayun, Neta Gadot, Ricky Shama, Stanley R. Rotman Department of Electrical and Computer Engineering, Ben-Gurion University of the Negev, P.O Box 653, 84105 Beer Sheva, Israel ABSTRACT Our research deals with developing metrics for quantizing the contrast of target to background in color images, and investigating the degree such contrast affects the detection of targets by human observers. When dealing with a gray scale image, the only parameter affecting the detection of targets is the luminance contrast; in color images, there are several parameters. This research examines these parameters and defines the importance of each one of them. Our system parameters quantifying color images are based on CIELAB space (brightness-color). We examine how each axis in this space affects the contrast of the targets. Keywords: Color, Target Acquisition 1. INTRODUCTION The human eye can differentiate up to one hundred shades of gray, while it can distinguish thousands of shades of color 1 . Therefore, the presentation of information in color should improve the detectability of targets in images. Moreover, the development of a model for quantitatively evaluating the contrast of targets in color images can help in assessing the performance of color detectors for military, industrial and medical needs. The subject of target detectability in images was discussed in the past by Ingman and Sheffer 1 , but only in gray-scale images. In recent years, due to the growing demand for the use of color images for target detection, there is a need to examine the human response to color images. Previous studies were done on the human visual system (HVS) determining the parameters for target recognition capability. Johnson 2 used a photo identification model for human observers. In his study, he investigated a method of analysis of electro-optical systems in spatial frequency space and defined essential criteria for the detection of military vehicles. Blackwell 3 added additional psychological effects to the target acquisition modeling. Van-Meeteren discovered 4 that the detection of non-uniform targets on non-uniform backgrounds is more difficult than for uniform targets on uniform backgrounds; he uses these results to build a predictive model based on experimental work. Sheffer et al. 5 studied in depth the work of Blackwell and Van-Mitrin and proposed a metric based on the mutual information difference between an image with a target and an image with no target. Driggers and Teaney were the first to extend the intensity contrast based on target acquisition models to a three dimensional color space (CIELAB) 6 . Target detectability is a multi-dimensional problem. There are many properties that affect target detectability, e.g. color, shape, and brightness. Geometrical properties such as the geometry of the target and the distance from the object also play an important role in an observer’s capability to detect a target. Hence, all those parameters must be taken by consideration in a completed model. This paper describes an adaptive contrast unit in color image represented in CIELAB space, which is designed to preserve contrast differences along all of its axes. The distance between the color and intensity of the target and the color and the intensity of the background should correspond to the contrast used in the Johnson model 2 . We follow in the direction of Ref. 6, presenting a new way to quantify the degree that the CIELAB space will answer our needs for the contrast metric. Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXIV, edited by Gerald C. Holst, Keith A. Krapels, Proc. of SPIE Vol. 8706, 87060N · © 2013 SPIE CCC code: 0277-786X/13/$18 · doi: 10.1117/12.2008849 Proc. of SPIE Vol. 8706 87060N-1 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

SPIE Proceedings [SPIE SPIE Defense, Security, and Sensing - Baltimore, Maryland, USA (Monday 29 April 2013)] Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXIV

Embed Size (px)

Citation preview

Producing a color target acquisition metric

Assaf Asbag, Racheli Hayun, Neta Gadot, Ricky Shama, Stanley R. Rotman

Department of Electrical and Computer Engineering, Ben-Gurion University of the Negev,

P.O Box 653, 84105 Beer Sheva, Israel

ABSTRACT

Our research deals with developing metrics for quantizing the contrast of target to background in color images, and

investigating the degree such contrast affects the detection of targets by human observers. When dealing with a gray

scale image, the only parameter affecting the detection of targets is the luminance contrast; in color images, there are

several parameters. This research examines these parameters and defines the importance of each one of them. Our system

parameters quantifying color images are based on CIELAB space (brightness-color). We examine how each axis in this

space affects the contrast of the targets.

Keywords: Color, Target Acquisition

1. INTRODUCTION

The human eye can differentiate up to one hundred shades of gray, while it can distinguish thousands of shades of color

1.

Therefore, the presentation of information in color should improve the detectability of targets in images. Moreover, the

development of a model for quantitatively evaluating the contrast of targets in color images can help in assessing the

performance of color detectors for military, industrial and medical needs.

The subject of target detectability in images was discussed in the past by Ingman and Sheffer1, but only in gray-scale

images. In recent years, due to the growing demand for the use of color images for target detection, there is a need to

examine the human response to color images.

Previous studies were done on the human visual system (HVS) determining the parameters for target recognition

capability. Johnson2 used a photo identification model for human observers. In his study, he investigated a method of

analysis of electro-optical systems in spatial frequency space and defined essential criteria for the detection of military

vehicles. Blackwell3 added additional psychological effects to the target acquisition modeling. Van-Meeteren

discovered4 that the detection of non-uniform targets on non-uniform backgrounds is more difficult than for uniform

targets on uniform backgrounds; he uses these results to build a predictive model based on experimental work. Sheffer et

al.5 studied in depth the work of Blackwell and Van-Mitrin and proposed a metric based on the mutual information

difference between an image with a target and an image with no target. Driggers and Teaney were the first to extend the

intensity contrast based on target acquisition models to a three dimensional color space (CIELAB)6.

Target detectability is a multi-dimensional problem. There are many properties that affect target detectability, e.g. color,

shape, and brightness. Geometrical properties such as the geometry of the target and the distance from the object also

play an important role in an observer’s capability to detect a target. Hence, all those parameters must be taken by

consideration in a completed model.

This paper describes an adaptive contrast unit in color image represented in CIELAB space, which is designed to

preserve contrast differences along all of its axes. The distance between the color and intensity of the target and the

color and the intensity of the background should correspond to the contrast used in the Johnson model2. We follow in

the direction of Ref. 6, presenting a new way to quantify the degree that the CIELAB space will answer our needs for the

contrast metric.

Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXIV, edited by Gerald C. Holst, Keith A. Krapels, Proc. of SPIE Vol. 8706, 87060N · © 2013 SPIE

CCC code: 0277-786X/13/$18 · doi: 10.1117/12.2008849

Proc. of SPIE Vol. 8706 87060N-1

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

2. IMAGE PARAMETERS AND COLOR SPACES

There are some important image parameters that need to be defined in this paper:

Contrast: Several definitions of contrast exist. In gray scale images, this parameter is defined by the difference between

the numerical luminance values of the pixels. In a color image, this definition is more complex because each pixel is

characterized by more than one property; it is the subject of this paper to achieve such a definition.

Intensity: The luminance brightness of a pixel in the image

Hue: The color shades from low (red) to high (blue) values.

Saturation: The mixture of white in the pixel; low values indicate a lack of color (gray) and high values indicate that the

dominant color is apparent.

HSI (Hue, Saturation, Intensity): The HSI space used to define the color of a pixel; it is but one of several different axes

to describe color.

RGB (red, green, and blue): The RGB model is the most common model in the field of digital images. According to this

model, each color is a combination of different levels of red, green and blue. These three colors are colors that are visible

to the human eye.

CIELAB: This model was initially developed in 1931 and corrected in 1976 to represent the perception of colors by the

human eye. The three-dimensional coordinate system is defined as a two dimensional plane for color, when the

orthogonal dimension is intensity. It should be noted that uniform hue or saturation are complex curves in this space.

This space is based on the photoreceptor of the human eye. Axis "a" represents the sensitive photoreceptor green and red

colors, axis "b" represents the sensitive photoreceptor yellow and blue, while the "L" axis represents the brightness

sensitive photoreceptor. The color is defined as:

and the hue is defined as:

Following the examination of the above spaces, we reached the conclusion that for our research the most compatible

space is CIELAB, due to its similarity to the response of the human eye.

3. CONTRAST UNIT AND CHANGING TARGET PARAMETERS

3.1 Contrast unit

We wish to scale our axes so that a unit change in any direction in our three-dimensional CIELAB space will be subject

to the same threshold of detection by a human observer. In addition, in our experiments, the modification of the contrast

must maintain a "natural environment" in relation to the background. In other words, adding / subtracting a contrast unit

to the target parameters adds no artificial parameters that make it easier for the viewer to identify the target. It maintains

a natural and gradual transition between the target and the background.

To produce images with a constant degree of contrast, we start with natural images with a small target present in each

image; see figure 1 for a sample of how the target and background histograms appear in CIELAB space. The target

appears mostly as a blurred Gaussian (see figure 2) We calculate the pixel distribution histograms of the target and local

(1)

(2)

Proc. of SPIE Vol. 8706 87060N-2

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

' ',

L notes-an artne noon ims1 lb MSS

ao

L

MSS

A

so

SOS

B

background areas; with those calculations we derived first-order statistics of the image, i.e. mean and standard deviation.

Second, we calculate the difference between the target and background mean in each axis. In these images this basically

consists of calculating the luminance of each target pixel relative to the average background. Third, we “alter the target",

changing the value in the target pixel. If previously the target had been three units above the background in the

luminance direction, we can now make it three units above (or below) the background in any of the three axes.

3.2 Changing Target Parameters

We change the target parameters by adding / subtracting contrast units to selected axes in a way that maintains a natural

transition between the background and the target, thereby creating a neutral environment to examine the effect of

changing contrast units in the examined axes. It is important to eliminate the influence of an axis which is not tested to

create a natural environment for examining the different parameters. We achieve this by setting the target mean of the

irrelevant axis to the background mean value. To ensure that these changes indeed maintain a neutral environment, we

add/subtract contrast units per pixel (see figure 3 and equation 3).

Figure 2: Transformation of target parameters

Figure 1: Histogram of an image in the CIELAB space. The background is given by the blue histogram; the target by the red.

(3)

Proc. of SPIE Vol. 8706 87060N-3

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

It

SAwtirN1. wfttle6Wo-sm. .rw. 6.r

1 rfsw!"

Q

1123

Q1we OM. aw1a. 1,..rw...rr.....

Figure 3: Process of changing target parameters

4. EXPERIMENTAL PROCEDURE

We performed a statistical experiment in order to examine the definition of the contrast unit as well as the effect of each

axis in CIELAB space and their optimal combination.

During the experiment we tested 60 observers’ ability to distinguish targets in images in which we change L, A, B

parameters of the target relative to its surroundings.

Each subject was presented 15 scenarios which repeated themselves in three different backgrounds for a total of 45 cases.

36 scenarios included four images and 9 questions had two images (for a total 45 questions). In each question, we

presented to the experimenter the original image with a marked target. Underneath, the target had been altered in

two/four different ways. The observer was asked to rank the targets in terms of their contrast. The experiment was

"forced choice" where each image had to be ranked from 1 to 4 (see figure 4).

Figure 4: GUI screen shots

Proc. of SPIE Vol. 8706 87060N-4

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

4.1 Experimental design

The subjects consisted of thirty men and thirty women. The experiment was conducted on three identical computers, in

the morning, in order to create identical conditions for all subjects.

Every subject received the same instructions before the start of the experiment (reading and audio). Before beginning the

experiment, each subject underwent a color blindness test. During the experiment, we measured the time it takes for an

subject to rank the pictures; we guaranteed the reliability of their ranking of the targets by our supervision of their

activities during the experiment.

To avoid the influence of external parameters, we displayed the questions in random order; furthermore, due to the

human brain tendency to rank the images from left to right, the questions and the images in every question was

counterbalanced between subjects. Subjects who ranked all the questions in the same way, took abnormal time to rank

the images, did not pass the color blindness test, or received an error message more than 5 times for improperly

executing the experiment were excluded from the experimental design; their ranking is not considered.

5. RESULTS AND CONCLUSIONS

5.1 Results

In order to allow qualitative analysis, we defined a weighted rank to analyze the results (Equation 4). The ranking is as

follows: the image which the Experimenter chose as first received 2 points, second place received one point, third place

received minus one point and the fourth received minus 2 points. To analyze the results for each question, the points

were summed to a weighted rank that served as the fundamental metric. The results are seen in figure 5 and tables1-2.

5.2 Conclusions

Brightness Axis: There is no difference in sensitivity between a positive / negative contrast regardless of the background

brightness axis. The L axis is more dominant than axes A and B. When changing the L axis along with another axis (A or

B), the change in the L axis is dependent on the background brightness regardless of the other changes in the second

axes.

‘A’ axis: There is an enhanced sensitivity for positive change (red over green) regardless of the color of the background.

In addition, Axis A is more dominant than axis B.

‘B’ axis: There is an enhanced sensitivity for negative change (blue over yellow). Axis B is the least dominant of the

three axes of the CIELAB space.

Changing color on both axes - A and B: The change of full contrast unit in both axes (A and B axes) is independent of

the background mean of the brightness axis. On the other hand, the change of half-contrast unit in the color axes (A and

B) is dependent on the background mean of the brightness axis: for a dark image there is a preferences for target with

bright colors (yellow and red), and for a bright image, there is a preferences for target with dark colors (green and blue).

(4)

Proc. of SPIE Vol. 8706 87060N-5

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

Three different pictures35 30

Quel

I 1 Quel30

25_ Que3

2520

d a1

20 L

15-0

15

1010

55

0 o2 3 4

Picture Num

Av of the pictures ranking

2 3 4

Picture Num

Changes First picture Second pictureFull change in L 26 34Full change in A 7 53Full change in B 59 1

First picture - negative change in the selected axisSecond picture - positive change in the selected axis

Changes First picture Second picture Third picture Fourth pictureHalf change in L and A -42 -39 5 43

Full change in L and A -51 19 -6 43

Half change in L and full in A -48 -34 35 49Full change in L and half in A -26 -8 15 33

Half change in L and B 33 -43 -6 16

Full change in L and B -15 22 -68 38Half change in L and full in B 31 27 -72 14

Full change in L and half in B -9 -10 -20 40Half change in A and B 32 -87 -33 70Full change in A and B 31 -59 -66 94Half change in A and full in B 80 -77 -75 70Full change in A and half in B -22 0 -61 83

First picture - negative change in both axesSecond picture - positive change in the first axis and negative change in the second oneThird picture - negative change in the first axis and positive change in the second oneFourth picture - positive change in both axes

Table 2 :

Table 1 :

Figure 5: Sample graph for a question

Proc. of SPIE Vol. 8706 87060N-6

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms

REFERENCES

[1] Ingman, D. and Sheffer, D., "Analyzing target recognition issues using the informational difference concept," Proc.

SPIE 3068, 136-147 (1997). [2] Johnson, J., “Analysis of image forming systems,” in Image Intensifier Symposium, AD 220160 (Warfare Electrical

Engineering Department, U.S. Army Research and Development Laboratories, Ft. Belvoir, Va.), 244–273 (1958). [3] Blackwell, H. R., "Contrast thresholds of the human eye", J. Opt. Soc. Am. 36(11), 624-643 (1946).

[4] Van Meeteren, A., "Characterization of task performance in viewing instruments,” J. Opt. Soc. Am. A 7(10), 2016-

2023 (1990).

[5] Sheffer, D., Kafri, A., Voskoboinik, A., Setter, P. and Norman, J. , "Use of the informational difference as a target

conspicuity measure," Proc. SPIE 5075, 150-160 (2003).

[6] Krapels, K. A., Jones, T., Driggers, R. G. and Teaney, B., "Target detection in color imagery: on the path to a color

target acquisition model," Proc. SPIE 5612, 295-303 (2004).

[7] Driggers, R. G., Krapels, K., Vollmerhausen, R., Warren, P., Scribner, D., Howard, G., Tsou, B. H. and Krebs, W.

K., "Target detection threshold in noisy color imagery," Proc. SPIE 4372, 162-168 (2001).

Proc. of SPIE Vol. 8706 87060N-7

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/19/2013 Terms of Use: http://spiedl.org/terms