Eye Movement Data Modeling Using a Genetic Algorithm Yun Zhang , Hong Fu, Zhen Liang, Xiaoyu Zhao, Zheru Chi, Dagan Feng and Xinbo Zhao Abstract—We present a computational model of human eye movements based on a genetic algorithm (GA). The model can generate elemental raw eye movement data in a four-second eye viewing window with a 25 Hz sampling rate. Based on the physiology and psychology characters of human vision system, the fitness function of the GA model is constructed by taking into consideration of five factors including the saliency map, short time memory, saccades distribution, Region of Interest (ROI) map, and a retina model. Our model can produce the scan path of a subject viewing an image, not just several fixations points or artificial ROI’s as in the other models. We have also developed both subjective and objective methods to evaluate the model by comparing its behavior with the real eye movement data collected from an eye tracker. Tested on 18 (9 h 2) images from both an obvious-object image group and a non-obvious-object image group, the subjective evaluations shows very close scores between the scan paths generated by the GA model and those real scan paths; for the objective evaluation, experimental results show that the distance between GA’s scan paths and human scan paths of the same image has no significant difference by a probability of 78.9% on average.

I. INTRODUCTION Due to an exponential increasing in the amount of images produced on a daily basis, image understanding plays the key role to the machine vision applications, such as image database management, image understanding and compression. Various computational models have been proposed to simulate human vision system [1 – 5, 11]. However, because of the big gap between the high level image interpretation and low level image features, those existing abstruse attention models’ perform inconsistently with different circumstances. Obviously, the breakthrough lies in a thoroughly understanding of human visual and perceptual mechanisms. Eye movement data may provide some insights on the behavior of human perception. Yun Zhang is with School of Computer Science, Northwestern Polytechnical University , Xi’an, Shaanxi, P. R. China ([email protected]) Hong Fu, Zhen Liang, Xiaoyu Zhao, Zheru Chi and Dagan Feng are with Center for Multimedia Signal Processing, Department of Electronic and Information Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong. (e-mail:[email protected], [email protected], [email protected], [email protected], [email protected] ). Dagan Feng is also with School of Information Technologies, The University of Sydney, NSW 2006, Australia. (e-mail: [email protected] ) Xinbo Zhao is with the Key Lab of Contemporary Design & Integrated Manu. Tech./ Northwestern Polytechnical University, Xi’an, Shaanxi, P.R. China ([email protected])

c 2009 IEEE 978-1-4244-2959-2/09/$25.00 

It is assumed that the fixation, saccade and smooth pursue can provide evidence of voluntary, overt visual attention. Pursuit movements are mainly observed visually tracking a moving target. Consequently, we have been studying and defining a computational genetic algorithm (GA) model of fixation and saccade mechanisms to simulate human eye movements in still image viewing. The motivation of eye movements’ research is to provide the unique insights into visual perception and human behavior, from which we can formulate the inner pre-attention and cognitive process and understand the human attention model. There has been very limited work reported in this area. Rutishauser and Koch [1] derived an eye movement model to find out which image features bias the underlying neuronal responses in conduction a search task. Their experiments were based on the simple synthetic stimulus which was well restricted in color, orientation and size to predict the modulation of firing rates of V4 and FEF neurons in searching tasks. Rybak [6] provided a visual perceptual model containing both the low-level and high-level subsystems. However, their model mainly simulated the recognition behavior of subjects on images (e.g. faces) invariantly with respect to shift, rotation and scale. Privitera and Stark [2] did a significant of developing the cognitive mechanism respectively for ten commonly image processing algorithms and evaluated the algorithmically detected region of interests (aROI’s) against human identified region of interests (hROI’s) using both subjective and objective methods. Similar to other three-second eye movement experiments, they clustered the raw eye movement data into seven fixations and showed the scan-path consistency by evaluating the distance between aROI’s and hROI’s. Jaimes et al. [3] and W. Zhang et al. [4] also investigated the computational eye movement models which were used in specific image understanding tasks such as object detection and image classification. The rest of this paper is organized as follows. Section II introduces experimental design and setting for collecting true eye movement data. A computational GA model for eye movements are discussed in Section III, in which we introduce a fitness function and present some GA simulated eye gaze data of four images from both the object group and non-object group. In section IV, we provide not only a subjective evaluation but also an objective evaluation. In the subjective method, we evaluate the performance of our GA model by using one-way ANOVA. In the objective evaluation, the Hausdorff distance is used to measure the similarity between the eye movement data produced by the GA model and those collected from an eye tracker. Finally, concluding remarks are mad in section V.

1038

II. EYE MOVEMENT RECORDING Commercial eye tracking products have grown in the last twenty years. The eye tracking products includes, SR Research’s Eyelinker, LC Technologies’ Eyegaze and EyeFollower, SMI’s iView, Tobii X120, etc. These products are usually of general type and not specially tailored to each visual task with different circumstances. Moreover, these products are generally quite expensive. Here we use the self-developed head mounted eye tracking system, Eye Secret, to collect eye gaze data (Fig. 1). The helmet surface data were collected by the ATOS measuring system, then the head gear was designed from these data and manufactured by the rapid prototyping technology (Dimension 3D Printer) for the best fitness and flexibility. The system computes the Point of Regard (POR) by distracting the corneal reflection and the center of the pupil from the image sequence. Eye Secret’s sampling rate is 25HZ, accuracy is about 1° visual angle, horizontal FOV (Field of View) is ±30° and vertical FOV ±25°.

Figure 1: The leˆ–ǣ–Š‡†‡•‹‰‘†‡Ž‘ˆEye Secret•›•–‡ǢŠ‡—’’‡” ”‹‰Š–ǣ–Š‡•‡––‹‰‘ˆ–Š‡‡›‡‘˜‡‡––‡•–•ǢŠ‡Ž‘™‡””‹‰Š–ǣ–Š‡ …ƒ’–—”‡†’—’‹Ž…‘”‡ƒŽ”‡ˆŽ‡…–‹‘‹ƒ‰‡—†‡”–Š‡ Ž‹‰Š–Ǥ

In order to estimate the eye movements during static image reading, the subject was seated in the front of the screen with his head secured onto an optometric chin-rest structure. The viewing distance was approximately 45 cm from the monitor with the screen dimensions 34cmh27cm with 0.264 mm pixel pitch, yielding a subtended visual angle of approximately 34°h42°. Eye movement traces of five subjects were recorded when they viewed a series of 18 images (Fig. 2) from both the obvious-object image group and the non-obvious-object image group (ten viewings each image). Each image is 635X420 pixels and 9 point calibrations are used before each testing.

(a) (b) ‹‰—”‡ʹǣȋƒȌ„˜‹‘—•Ǧ‘„Œ‡…–‹ƒ‰‡‰”‘—’ȋ„Ȍ‘Ǧ‘„˜‹‘—•Ǧ‘„Œ‡…–‹ƒ‰‡ ‰”‘—’Ǥ

III. COMPUTATIONAL MODEL FOR EYE MOVEMENT MODELING A. Theory The photoreceptor array in the retina is anisotropic and the effective receptor density falls dramatically within one degree of the central fovea. When the observers either hold their gazes at a stationary point (fixations) or move them quickly between those fixations (saccades), a great deal of information could gather. The saccades are rapid, ballistic movements that reach a velocity of over 500 degrees/second. From less than one degree to over 90 degrees, the duration of the saccades are typically completed in less than 50 ms which can be fully recorded by our eye tracker Eye Secret. In general, approximately three to four saccadic eye movements are made per second. Most of the studies reported on eye tracking are concerned of the 7 ~ 11 fixation sequence in a four-second tracking experiment from the filtered and clustered raw eye movement data (Fig.3).Thus, their computational frameworks generally produce the same kind of simplified analytical fixation sequence or areas but not the sampling eye movement data.

(a) (b) Figure 3: (a) 100 simulated eye gaze points produced by a GA-based model; (b) The algorithm-based 6-fixation sequence which was clustered from the raw eye movement data.

B. GA model Genetic algorithms (GA) have been used to solve nonlinear optimization by simulating evolution, i.e., the survival of the fittest strategy. Our GA model is try to simulate the scan paths of a single subject when reading different types of images, say, non-obvious object and obvious object group. The key issues here are the selection of chromosome representation and the design of fitness functions. The GA model will produce a subject’s four-second image viewing data at a 40~50 ms interval, that is, a one hundred eye gaze position sequence ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻǡ ݅ ൌ ͳǡʹǡ ‫ ͲͲͳ ڮ‬. A simple binary coded chromosome is adopted to encode x and y. The fitness function is constructed to imitate the low-level subsystem of the subject which performs a fovea-like transformation by detecting the primary features of an image. In doing so, five factors were taken into consideration: 1) Saliency Map Saliency map is a topographically arranged map that represents visual saliency of a corresponding visual scene [12]. At a given location, ܵ‫݌ܽܯ‬ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ is determined primarily by how different it is from the surrounding color, orientation, motion, depth etc, which finally integrates the normalized information from the individual feature maps into

2009 IEEE Congress on Evolutionary Computation (CEC 2009)

1039

one global measure of conspicuity [15]. During the pre-attention stage, low level image features are parallel processed by the human visual system to form a master activation map for the saccades control [16]. Saliency map is plausible suggestion to produce the saccades by a reasonable spatial scale (usual coarser than the detailed original image), so in the fitness function, ܵ‫݌ܽܯ‬ௌ௔௖௖௔ௗ௘௦ ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ will be introduced to denote the saliency value calculated at a saccades-related resolution scale. 2) Short time memory The attention mechanism has the short time memory to store the last m fixated locations. If the item with higher ƒ’ሺš୧ ǡ ›୧ ሻ is stored in the memory, the next highest value is chosen in each iteration [1]. 3) Saccades distribution According to E. M. Van Loon [8], human saccades follows a Gamma distribution with a mean of 5.3f8.0(f standard deviation). In our experiment, the viewing distance was approximately 45cm and the monitor’s pixel pitch is 0.264mm, so that the average span between each fixation should be about 14 pixels. 4) ROI Map Generally, a four-second image viewing window contains about 7 ROI’s which guide the distribution of the set of delicate oculomotor points. Known as the input of controlling mechanism for covert selective attention, saliency maps of the subjects’ viewing image are used again to form the ROI heat map. 5) Retina Model According to Rybak’s attention model [5], there is a decrement in resolution from the fovea to the retinal periphery in the cortical map of retina, that is to say the distribution of the fixation is not due to the size of ROI’s but the complexity of ROI’s. As a result, we use Ɂ୰ୣୱ୭୪୳୲୧୭୬ to control the span of each saccade (1°~40°) to be in the ROI’s. (micro-saccades) or between ROI’s. . By considering the five factors mentioned above, we define a fitness function for the GA model as ݂݅‫ݏݏ݁݊ݐ‬ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ  ൌ ܵ‫݌ܽܯ‬ௌ௔௖௖௔ௗ௘௦ ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻȂ

variations of different subjects). Ɂ୰ୣୱ୭୪୳୲୧୭୬ will produce a impulse response to compel the next fixation jump to different ROI regions. Constant ݇ sets a weighting relative to ܵ‫݌ܽܯ‬ௌ௔௖௖௔ௗ௘௦ ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ, here we set ݇ ൌ ͳ. Concretely, each eye gaze position ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻǡ ݅ ൌ ͳǡʹǡ ‫ͲͲͳ ڮ‬ is iteratively generated by running the GA model 100 times. The initial population of each gaze position ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ is randomly selected. Withሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ, both ܵ‫݌ܽܯ‬ௌ௔௖௖௔ௗ௘௦ ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ σ ܴܱ‫݌ܽܯܫ‬ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ will be update for next and point ሺ‫ݔ‬௜ାଵ ǡ ‫ݕ‬௜ାଵ ሻ . Due to the memory and retina window, σ  ƒ’ counts the gaze points successively falling in the same region. If it reach ρϐ୧୶ୟ୲୧୭୬ ,ߜ௥௘௦௢௟௨௧௜௢௡ will be activated and then  ƒ’ and ρϐ୧୶ୟ୲୧୭୬ will be refreshed for the next point. Some of the parameter controlling the GA model is given in Table I. TABLE I SOME PARAMETERS FOR TRAINING THE GA MODEL Probability Probability Population No. of Selection of the one of binary Size Generations Function point mutation crossover Normalized 100 15 geometric 0.6 0.05 selection

We run the GA model on 18 (9h2) images from two image groups mentioned in section II, some of the results are shown in Fig 4. In the second column of Fig. 4, we also generate the corresponding heat map which is conventionally used to inspect the visual attention mechanism. The last column in Fig. 4, shown is the final GA created 100 fovea transformation points during image viewing. Combined with the heat map, we can clearly see that the GA results cover the significant contents of image and have a similar distribution as human data. The concrete evaluation is given in section IV.

݇ ȉ ߜ௥௘௦௢௟௨௧௜௢௡ ൫σ ܴܱ‫݌ܽܯܫ‬ሺ‫ݔ‬௜ ǡ ‫ݕ‬௜ ሻ െ ߤ௙௜௫௔௧௜௢௡ ൯ȋͳȌ In the fitness function Eq.(1) the first item ƒ’ୗୟୡୡୟୢୣୱ ሺš୧ ǡ ›୧ ሻ is the saliency map modulated by a saccades distribution. In order to generate more realistic eye movement data, ƒ’ୗୟୡୡୟୢୣୱ ሺš୧ ǡ ›୧ ሻ controls the span between each pair of fixation. ƒ’ୗୟୡୡୟୢୣୱ provides a saliency map at median 5.3esaccade level which ensures that the distance between successive fixations are averagely about 14 pixels. The setting of this parameter is based on our experimental setting described in Section ,, The second item Ɂ୰ୣୱ୭୪୳୲୧୭୬ ሺσ  ƒ’ሺš୧ ǡ ›୧ ሻ െ ρϐ୧୶ୟ୲୧୭୬ ሻ is the control of each saccade’s span, where, σ  ƒ’ሺš୧ ǡ ›୧ ሻ is the count of the micro-saccades continuously within the same ROI region, if the count number reaches the ρϐ୧୶ୟ୲୧୭୬ (the experience value is 4 ~ 9, due to

1040

(a)

(b)

(c)

‹‰—”‡ Ͷǣ ‘‡  •‹—Žƒ–‡† ‡›‡ ‰ƒœ‡ †ƒ–ƒǤ Š‡ ˆ‹”•– –™‘ ‹ƒ‰‡• ƒ”‡ ˆ”‘ –Š‡ ‘„˜‹‘—•Ǧ‘„Œ‡…– ‹ƒ‰‡ ‰”‘—’ ƒ† –Š‡ •‡…‘† –™‘ ‹ƒ‰‡• ƒ”‡ ˆ”‘–Š‡‘Ǧ‘„˜‹‘—•Ǧ‘„Œ‡…–‹ƒ‰‡‰”‘—’ǤȋƒȌŠ‡‘”‹‰‹ƒŽ‹ƒ‰‡•Ǣȋ„Ȍ Š‡ ‹ƒ‰‡• ™‹–Š Š‡ƒ–Ǧƒ’•Ǣ ȋ…Ȍ Š‡ ‹ƒ‰‡• ™‹–Š Š‡ƒ– ƒ’• ƒ†  •‹—Žƒ–‡†‡›‡‰ƒœ‡†ƒ–ƒǤ

2009 IEEE Congress on Evolutionary Computation (CEC 2009)

SUBJECTIVE EVALUAT TION OF THE EYE MOVEMENT DATA GENERAT TED BY THE GA MODEL OBVIOUS-OBJECT IMAGE

IV. EVALUATION OF EYE MOVEMENT DATTA PRODUCED BY THE MODEL

In order to assess the similarity betweeen GA viewing patterns and those of humans perceptually aand quantitatively, subjective evaluation and objective evaluation are investigated here.

GROUP

IMAGE NO 1 2 3 4

A. Subjective Evaluation Qualitative or subjective evaluation proovides an insight into the effectiveness of the proposed modell. We mixed three groups of human eye scan paths and threee groups of GA’s scan paths. The sequences are redrawn on tthe corresponding images and replayed to the invited review wers (six subjects) who may and may not be familiar with eye ttracking research. The reviewers are exposed to 18 images wiith replay of gaze point sequences in viewing images. Each of the reviewers is asked to grade on those scan paths without kknowledge of how many are genuine eye gaze data and how m many are from the GA model. Scores 5 to 1 represents five different grades compared with real human eye gaze data: perfect, good, final results are acceptable, medium, and bad. The fi summarized in Table II. Table II shows the statistical results of thhe scores for both the human and GA scan paths on all the 18 iimages. The same results are also visualized in Fig. 5 and F Fig. 6. Generally speaking, the result shows very close scorees for both human scan paths and GA produced scan paths. Thiis is evidenced by the average scores of human and GA scan ppaths on both the obvious-object image group and non-obviouus-object images. We noticed that the average score of the GA A mode is slightly higher than that of human scan paths. Thiss may be because the three groups of human data include soome “overflowing gaze points” when test subjects look out off the pictures and some “singular gaze points” when subjeccts feel tired and bored of repeated presentations of the same image. The subjective evaluation indicates that the invitted reviewers can hardly tell the differences between the sim mulated scan paths and genuine ones. This suggests that thee GA model can produce realistic scan paths. Quantitativve evaluation is discussed in the next section.

5 6 7 8 9 AVERAGE

6.0

IMAGE GROUP

HUMAN

GA SCAN

SCAN PATH

PATH

2.7±0.6 3.7±0.4 3.1±1.3 2.7±1.5 3.2±1.8 3.5±2.0 3.6±1.7 4.1±1.4 2.9±1.5 3.3±0.5

NON-OBVIOUS-OBJECT

3.8±1.2 2.5±0.1 4.9±0.2 2.6±1.8 2.3±0.7 2.9±1.6 4.3±1.1 2.9±0.7 4.6±0.7 3.4±1.0

Human scan path

GA SCAN

HUMAN SCAN PATH

3.5±0.9 4.0±1.2 4.4±0.8 3.2±1.8 3.0±2.1 4.0±1.2 2.9±1.5 2.7±0.2 4.5±0.9 3.6±0.7

PATH

4.5±0.4 3.9±1.3 4.8±0.2 2.9±1.9 5.0±0.4 3.6±1.8 3.8±1.3 3.4±0.5 4.3±0.3 4.0±0.7

GA scan path

5.0 4.0 3.0 2.0 1.0 0.0

 —•Ǧ‘„Œ‡…–‹ƒ‰‡‰”‘—’Ǥ ‹‰—”‡ͷǣ—„Œ‡…–‡˜ƒŽ—ƒ–‹‘•‘–Š‡‘„˜‹‘—

6.0

Human scan path

GA scan path

5.0 4.0 3.0 2.0

B. Objective Evaluation We adopt the Hausdorff distance [10, 133] to represent the similarity of the different scan paths from saame subject or the similarity between the scan paths generated by the GA model and those collected from subjects. The Hauusdorff distance is defined as

1.0 0.0

‫݌‬௫‫א‬஻ ݀ሺ‫ݔ‬ǡ ‫ܣ‬ሻ (2) ‫ܪܦ‬ሺ‫ܣ‬ǡ ‫ܤ‬ሻ ൌ ‫ݔܽܯ‬ሺ‫݌ݑݏ‬௫‫א‬஺ ݀ሺ‫ݔ‬ǡ ‫ܤ‬ሻǡ ‫݌ݑݏ‬ ™Š‡”‡ ǡ  ƒ”‡ …Ž‘•‡† •‡–• ƒ†݀ሺ‫ݔ‬ǡ ‫ܤ‬ሻ–Š‡ …Žƒ••‹…ƒŽ —…Ž‹†‡ƒ†‹•–ƒ…‡ˆ”‘’‘‹–š–‘‘„Œ‡…–Ǥ

 „˜‹‘—•Ǧ‘„Œ‡…–‹ƒ‰‡‰”‘—’Ǥ ‹‰—”‡͸ǣ—„Œ‡…–‡˜ƒŽ—ƒ–‹‘•‘–Š‡‘Ǧ‘„

TABLE II

2009 IEEE Congress on Evolutionary Computation (CEC 2009)

1041

During the human experiment ,each of the image (j) has been viewed for 10 times by each subject, and each scan-path ୨ can be denoted as : …ƒ୧ ǡ ‹ ൌ ͳǡʹǡ ǥ ǡͳͲǢ Œ ൌ ͳǡʹǡ ǥ ͳͺ; where stands for the ݅‫ ݄ݐ‬viewing , stands for 18 images; Similarly, the scan-path produced by GA can be denoted as :

…ƒ୨ ǡ Œ ൌ ͳǡʹǡ ǥ ͳͺ; where stands for 18 images. So we have:

1)

The distance among 10 times of viewings (the j-th image) by one subject: ௝



‫ܪܦ‬൫‫݊ܽܿܵܪ‬௔ ǡ ‫݊ܽܿܵܪ‬௕ ൯ ൌ ௝



‫ݔܽܯ‬ሺ‫݌ݑݏ‬௫‫א‬ுௌ௖௔௡ೕ ݀൫‫ݔ‬ǡ ‫݊ܽܿܵܪ‬௕ ൯ǡ ‫݌ݑݏ‬௫‫א‬ுௌ௖௔௡ೕ ݀൫‫ݔ‬ǡ ‫݊ܽܿܵܪ‬௔ ൯ሻ ೌ



(3)

 ‹‰—”‡ ͹ǣ  Š‡ ‰ƒœ‡ ’‘‹–• ‘ˆ –Š‡  ‘†‡Ž ƒ† ‘‡ ‘ˆ •—„Œ‡…–ǯ• –‡ ˜‹‡™‹‰•ǤŠ‡”‡†’‘‹–•ƒ”‡–Š‘•‡‰‡‡”ƒ–‡†„›–Š‡ ‘†‡Žƒ†–Š‡ ›‡ŽŽ‘™’‘‹–•ƒ”‡–Š‘•‡…‘ŽŽ‡…–‡†ˆ”‘–Š‡‡›‡–”ƒ…‡”ȋͳͲͲ‰ƒœ‡’‘‹–• ‹–‘–ƒŽȌǤ

where ݆ stands for the ݆‫ ݄ݐ‬of 18 images, ܽǡ ܾ ൌ ௝ ͳǡʹǡ ǥ ǡͳͲ,‫݊ܽܿܵܪ‬௔ is the ܽ‫ ݄ݐ‬human scan-path of the ݆‫݄ݐ‬ image ; 2)

1)

The Hausdorff distance between each pair of one subject’s 10 viewings of the image No. 6 from the obvious-object image group is shown in Table III.

TABLE III THE HAUSDORFF DISTANCE MEASURE BETWEEN EACH PAIR OF ONE SUBJECT’S 10 VIEWINGS OF IMAGE NO. 6

The distance among GA and 10 times of viewings (the j-th image) by one subject:





‫ݔܽܯ‬ሺ‫݌ݑݏ‬௫‫ீא‬஺ௌ௖௔௡ೕ ǡ ݀൫‫ݔ‬ǡ ‫݊ܽܿܵܣܩ‬௝ ǡ ൯ǡ ‫݌ݑݏ‬௫‫א‬ுௌ௖௔௡ೕ ݀൫‫ݔ‬ǡ ‫݊ܽܿܵܪ‬௔ ൯ሻ ೌ



ࡰࡴ൫‫݊ܽܿܵܪ‬௔ ǡ ‫݊ܽܿܵܪ‬௕ ൯ ǡ ݆ ൌ ͸in Object Group (Unit: Pixel)



‫ܪܦ‬൫‫݊ܽܿܵܣܩ‬௝ ǡ ǡ ‫݊ܽܿܵܪ‬௔ ൯= b a

1

2

3

4

5

6

7

8

9

10

(4)

1

0

153

219

165

262

131

214

193

227

222

where ‫ ݊ܽܿܵܣܩ‬௝ is the GA generated scan-path of the ݆‫݄ݐ‬

2

---

0

179

235

130

229

270

276

131

123

3

---

---

0

271

243

307

233

280

223

259

4

---

---

---

0

183

161

284

122

170

289

5

---

---

---

---

0

222

308

255

169

219

6

---

---

---

---

---

0

222

116

208

293

7

---

---

---

---

---

0

264

266

333

8

---

---

---

---

---

---

---

0

246

281

9

---

---

---

---

---

---

---

---

0

232

10

---

---

---

---

---

---

---

---

---

0

image, other items are same as Eq. (3).

C. Objective Evaluation of a Sample Image In this part, we will take Image No. 6 from the obvious-object image group to explain each step of the objective evaluation. In Fig. 7, the gaze points generated by the GA model and those by a subject for image No. 6 are shown in Fig. 7.

2)

1042

The Hausdorff distance measure between each GA model generated scan path and the subject’s viewing of the image No.6 from the obvious-object image group is shown in Table IV.

2009 IEEE Congress on Evolutionary Computation (CEC 2009)

TABLE IV THE HAUSDORFF DISTANCE MEASURE BETWEEN THE SCAN-PATH FROM GA MODEL AND SUBJECT’S VIEWING OF IMAGE NO. 6 ௝ ࡰࡴ൫‫݊ܽܿܵܣܩ‬௝ ǡ ‫݊ܽܿܵܪ‬௔ ൯ǡ ݆

ൌ ͸in Object Group

VI and VII, respectively. TABLE VI THE ANOVA TEST RESULT ON THE IMAGES FROM THE OBVIOUS-OBJECT IMAGE GROUP

(Unit: Pixel) Image No.

F

Prob > F

24

1

0.2302

0.6334

25

2

0.0542

0.8167

3

549

3

0.1987

0.6576

4

25

4

0.0572

0.8118

5

23

5

0.1036

0.7489

6

25

6

0.0659

0.7984

7

350

7

0.1606

0.6902

8

22

8

0.195

0.6606

9

363

9

0.0974

0.7563

10

345



Scan-path NO.

‫ܪܦ‬൫‫݊ܽܿܵܣܩ‬௝ ǡ ‫݊ܽܿܵܪ‬௔ ൯

1 2

TABLE VII THE ANOVA TEST RESULT ON THE IMAGES FROM THE NON-OBVIOUS-OBJECT IMAGE GROUP

 3)

One-way ANOVA results is shown in Table V (refer to Appendix for the details). TABLE V ONE-WAY ANOVA RESULT ANOVA Table

Source

SS

df

MS

F

Prob > F

Groups

451.0904

1

451.0904

0.050853

0.82245

Error

470132.9

53

Total

470584

54

8870.431

In Table V, the SS stands for the sum of squares due to each source; df stands for the degrees of freedom associated with each source; MS stands for the mean squares for each source, which can be calculated by SS/df; F stands for the ratio of the mean squares; Prob is the p-value derived from the cumulative distribution function (cdf) of F. We obtain an F-value of 0.051 which is much smaller than the cut-off value of ‫ܨ‬଴Ǥ଴ହ ሺͳǡͷ͵ሻ ൌ ͶǤͲʹ for F-distribution at 1 and 53 degrees of freedom and 95% confidence. Therefore, we accept the null hypothesis that there is no difference in ࣆࡰࡴቀࡴࡿࢉࢇ࢔࢐ ǡࡴࡿࢉࢇ࢔࢐ ቁ and ࣆࡰࡴቀࡳ࡭ࡿࢉࢇ࢔࢐ ǡࡴࡿࢉࢇ࢔࢐ ቁ with the ࢇ





probability P = 82.2%. That is to say, the distance between GA scan-path and human scan-path of one image is affected by the same one factor by a probability of 82.2%. The same one factor here can be considered as the same subject who is simulated by our GA model.

4)

One-way ANOVA results for 18 (9 × 2) images from the obvious-object image group and the non-obvious-object image group are shown in Tables

Image No.

F

Prob > F

1

0.2314

0.7494

2

0.0555

0.9328

3

0.1999

0.7737

4

0.0585

0.9279

5

0.1048

0.8649

6

0.0671

0.9145

7

0.1619

0.8062

8

0.1962

0.7767

9

0.0986

0.8723

From the ANOVA results for all the images we can see that averagely the P-value is about 78.85%, which is similar to the subject test result. Obviously the GA model considerably agrees with the human data.

V. CONCLUSION AND FUTURE WORK In this paper, we developed a genetic algorithm (GA) model to simulate human eye movements based on a fitness function considering five factors including the saliency map short time memory, saccades distribution, Region of Interest (ROI) map, and a retina model. Supported by a self-developed eye tracker Eye Secret, we recorded five subjects’ scan paths of viewing 18 images from both an obvious-object image group and a non-obvious-object image group. The scan paths generated by the GA model and those collected by the eye tracker were used for both subjective and objective evaluation of the performance of our GA model Tested on 18 (9 × 2) images, the subjective evaluation shows very close scores between the human scan paths and those

2009 IEEE Congress on Evolutionary Computation (CEC 2009)

1043

generated by the GA model; for the objective evaluation, the results show that the distance between GA scan paths and human scan paths of the same image have no significant difference by a probability of 78.85% on average. Although promising results have been achieved on modeling eye gaze data using a GA model, there are still a lot to improve the present GA frame to simulate different subjects’ viewing patterns on different images. The research results obtained so far from a good foundation for future research on image understanding and image content classification based on eye tracking data. ACKNOWLEDGMENTS

[6]

[7]

[9] [10]

[11]

APPENDIX One-way analysis of variance (ANOVA) tests is to determine if one given factor has a significant effect on the group under study. To evaluate the similarity between the scan paths generated by the GA eye movement model and real people’s visual perception of the same image, we have generated two sets of similarity values or the treatments: ࡰࡴ൫ࡴࡿࢉࢇ࢔࢐ࢇǡ ࡴࡿࢉࢇ࢔࢐࢈ ൯ andࡰࡴ൫ࡳ࡭ࡿࢉࢇ࢔࢐ ǡ ࡴࡿࢉࢇ࢔࢐ࢇ ൯ . So the null hypothesis H0 is: ࣆࡰࡴቀࡴࡿࢉࢇ࢔࢐ ǡࡴࡿࢉࢇ࢔࢐ ቁ ൌ ࣆࡰࡴቀࡳ࡭ࡿࢉࢇ࢔࢐ ǡࡴࡿࢉࢇ࢔࢐ ቁ ࢇ



[5]

[8]

The work reported in this paper is substantially supported by a research grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No.: Polyu 5141/07E) and a Polyu research grant (Project No.: 1-BB9V).



[4]

[12]

[13] [14] [15] [16]

Proceedings of SPIE Human Vision and Electronic Imaging VI, CA(2001), pp. 373--384 W. Zhang, H. Yang, D. Samaras, and G. Zelinsky, “A Computational Model of Eye Movements during Object Class Detection,” In Advances in Neural Information Processing (NIPS) 18, 2005, Vancouver, Canada, pp. 1609-1616. I.A. Rybak, V.I. Gusakova, A.V. Golovan, L.N. Podladchikova, and N.A. Shevtsova, “A model of attention-guided visual perception and recognition,” Vision Research, vol. 38 (1998), pp. 2387-2400. Y. Zhang, X. Zhao, R. Zhao, Y. Zhou, and X. Zou, “EyeSecret: an inexpensive but high performance auto-calibration eye tracker,” Proceedings of the 2008 symposium on Eye tracking research & applications, pp. 103-106. C. Koch and S. Ullman, “Shifts in selective visual attention: towards the underlying neural circuitry,” Human Neurobiology 4:219-227 (1985). E. M. Van Loon, I. Th. C. Hooge, and A. V. Van den Berg , “The Timing of Sequences of Saccades in Visual Search,” Proceedings: Biological Sciences, Vol. 269, No. 1500 (Aug. 7, 2002), pp. 1571-1579 L. Fisher, Fixed Effects Analysis of Variance, New York: Academic Press, 1978. D.P. Huttenlocher, G.A. Klanderman, and W.A. Rucklidge, “Comparing Images Using the Hausdorff Distance,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15, no. 9, pp. 850-863, September, 1993. H. Fu, Z. Chi, and D. Feng, “Attention-Driven Image Interpretation with Application to Image Retrieval,” Pattern Recognition, vol. 39, Issue 9, September, 2006, Pages 1604-1621. L. Itti, C. Koch, and E. Niebur, “A Model of Saliency-Based Visual Attention for Rapid Scene Analysis,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 11, November. 1998, pp. 1254-1259. J.F. Hangouet, “Computation of the Hausdorff Distance Between plane Vector Polylines,” ACSM/ASPRS Annual Convention & Exposition Technical Papers. Bethesda: ACSM/ASPRS, 1995, 4:1-10. C.R. Houck, J. Joines, and M. Kay, "A Genetic Algorithm for Function Optimization: A Matlab Implementation," NCSU-IE TR 95-09, 1995. D. Walther and C. Koch (2006), “Modeling attention to salient proto-objects,” Neural Networks, 19, 1395-1407. J. M. Findlay, “Saccade Target Selection During Visual Search,” Vision Research, Vol 37, Issue 5, March 1997, pp. 617-631.

and the alternative hypothesis H1 is: not H0 ,which means that the two levels of the factor are significantly different from each other. There are altogether 18 images and we take one-way ANOVA for each of them with the cut off value: ஑ ሺ• െ ͳǡ  െ •ሻ ൌ ‫ܨ‬଴Ǥ଴ହ ሺͳǡͷ͵ሻ ൌ ͶǤͲʹ where Ƚ ൌ ͲǤͲͷ , which is common significance levels; s = 2 (two groups); n = (45 +10) = 54, which is the sum of the sizes of two groups. So, ሺ• െ ͳǡ  െ •ሻ = (1, 53) are the degrees of freedom for both between and within groups. Calculate the ratio of the mean squares F, and then compare it with ‫ܨ‬଴Ǥ଴ହ ሺͳǡͷ͵ሻ , if ‫ ܨ‬൏ ‫ܨ‬଴Ǥ଴ହ ሺͳǡͷ͵ሻ , we accepted H0 ,otherwise, reject H0 . The p-value given in one-way ANOVA for each of the 18 images describes the probability of the outcome under the null hypothesis. REFERENCE: [1] [2]

[3]

1044

U. Rutishauser and C .Koch, “Probabilistic modeling of eye movement data during conjunction search via feature-based attention,” Journal of Vision, Apil, 2007, vol 7(6):5, 1-20. C.M. Privitera and L.W. Stark, “Algorithms for Defining Visual Regions-of-Interest: Comparison with Eye Fixation,” IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 22, pp. 970-982, 2000. A. Jaimes , J. B. Pelz , J. S. Babcock, and S. Chang, “Using Human Observers' Eye Movements in Automatic Image Classifiers,”

2009 IEEE Congress on Evolutionary Computation (CEC 2009)

Eye Movement Data Modeling Using a Genetic ... - Semantic Scholar

Dagan Feng is also with School of Information Technologies, The. University of .... effective receptor density falls dramatically within one degree of the central ...

1MB Sizes 1 Downloads 313 Views

Recommend Documents

Eye Movement Data Modeling Using a Genetic ... - Semantic Scholar
Yun Zhang is with School of Computer Science, Northwestern. Polytechnical University ... Manu. Tech./ Northwestern Polytechnical University, Xi'an, Shaanxi, P.R. .... movements that reach a velocity of over 500 degrees/second. From less ...

Lightpath Protection using Genetic Algorithm ... - Semantic Scholar
connectivity between two nodes in the network following a failure by mapping ... applications and high speed computer networks because of huge bandwidth of ...

Lightpath Protection using Genetic Algorithm ... - Semantic Scholar
virtual topology onto the physical topology so as to minimize the failure ... applications and high speed computer networks because of huge bandwidth of optical ...

Acoustic Modeling Using Exponential Families - Semantic Scholar
For general exponential models, there is no analytic solution for maximizing L(θ) and we use gradient based numerical op- timization methods. This requires us ...

Dynamic Channel Allocation Using a Genetic ... - Semantic Scholar
categorised in terms of a Channel Allocation Matrix. A novel. Dynamic ... same category in the proposed Channel Allocation Matrix. .... Priority decrease. 3 5 2 11 9. Fig. 5. An example of an individual's string representation. The fitness value is e

Dynamic Channel Allocation Using a Genetic ... - Semantic Scholar
Internet users that do not enjoy fast access networks. Broadband fixed wireless access (BFWA) networks can be rapidly deployed and provide data rates that current Internet users demand. Typical BFWA network components are the ..... operating in the 5

1 feature subset selection using a genetic algorithm - Semantic Scholar
Department of Computer Science. 226 Atanaso Hall. Iowa State ...... He holds a B.S. in Computer Science from Sogang University (Seoul, Korea), and an M.S. in ...

Acoustic Modeling Using Exponential Families - Semantic Scholar
The training set consisted of about 500hrs of audio. The LDA feature vectors were obtained by first com- puting 13 ... Again, Hybrid Monte Carlo seems to be best among .... 6th European Conference on Speech Communication and Technol-.

Acoustic Modeling Using Exponential Families - Semantic Scholar
shape parameter), Chi-square, and of course the diagonal and full covariance ..... Again, Hybrid Monte Carlo seems to be best among the Markov Chain .... Carlo methods,” Dept. of Computer Science, University of Toronto,. Tech. Rep.

Likelihood-based Data Squashing: A Modeling ... - Semantic Scholar
Sep 28, 1999 - performing a random access than is the main memory of a computer ..... identi es customers who have switched to another long-distance carrier ...

Genetic diversity in palmyrah genotypes using ... - Semantic Scholar
Email: [email protected] collection of neera and .... amplified the products in all the templates used. The ... DNA as template, 1 unit(0.3 µl) of Taq DNA.

A Generalized Data Detection Scheme Using ... - Semantic Scholar
Oct 18, 2009 - We evaluated the performance of the proposed method by retrieving a real data series from a perpendicular magnetic recording channel, and obtained a bit-error rate of approximately 10 3. For projective geometry–low-density parity-che

Quantifying Organismal Complexity using a ... - Semantic Scholar
Feb 14, 2007 - stomatitis virus, and to illustrate the consistency of our approach and its applicability. Conclusions/Significance. Because. Darwinian evolution ...

Data centric modeling of environmental sensor ... - Semantic Scholar
Email: {rdantu, kaja}@cs.unt.edu, [email protected], [email protected] ... Data mining techniques [9][10][21] on these datasets are useful to correlate ecological ...

Data centric modeling of environmental sensor ... - Semantic Scholar
and analysis of data centric algorithms for the collaborative retrieval .... Fig. 3 describes the error rate induced in. 10. 20. 30. 40. 50. 60. 70. 80. 90. 100. 0. 0.5. 1.

using rapd markers - Semantic Scholar
based on this, cluster analysis was done using minimum variance algorithm. Cluster analysis showed two major groups. Each sub-group was characterized ...