`
`IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 20, NO. 11, NOVEMBER 1998
`
`Short Papers
`
`A Model of Saliency-Based Visual Attention
`for Rapid Scene Analysis
`
`Laurent Itti, Christof Koch, and Ernst Niebur
`
`Abstract—A visual attention system, inspired by the behavior and the
`neuronal architecture of the early primate visual system, is presented.
`Multiscale image features are combined into a single topographical
`saliency map. A dynamical neural network then selects attended
`locations in order of decreasing saliency. The system breaks down the
`complex problem of scene understanding by rapidly selecting, in a
`computationally efficient manner, conspicuous locations to be analyzed
`in detail.
`
`Index Terms—Visual attention, scene analysis, feature extraction,
`target detection, visual search.
`
`———————— F ————————
`
`1 INTRODUCTION
`PRIMATES have a remarkable ability to interpret complex scenes in
`real time, despite the limited speed of the neuronal hardware avail-
`able for such tasks. Intermediate and higher visual processes appear
`to select a subset of the available sensory information before further
`processing [1], most likely to reduce the complexity of scene analysis
`[2]. This selection appears to be implemented in the form of a spa-
`tially circumscribed region of the visual field, the so-called “focus of
`attention,” which scans the scene both in a rapid, bottom-up, sali-
`ency-driven, and task-independent manner as well as in a slower,
`top-down, volition-controlled, and task-dependent manner [2].
`Models of attention include “dynamic routing” models, in
`which information from only a small region of the visual field can
`progress through the cortical visual hierarchy. The attended region
`is selected through dynamic modifications of cortical connectivity
`or through the establishment of specific temporal patterns of ac-
`tivity, under both top-down (task-dependent) and bottom-up
`(scene-dependent) control [3], [2], [1].
`The model used here (Fig. 1) builds on a second biologically-
`plausible architecture, proposed by Koch and Ullman [4] and at
`the basis of several models [5], [6]. It is related to the so-called
`“feature integration theory,” explaining human visual search
`strategies [7]. Visual input is first decomposed into a set of topo-
`graphic feature maps. Different spatial locations then compete for
`saliency within each map, such that only locations which locally
`stand out from their surround can persist. All feature maps feed, in
`a purely bottom-up manner, into a master “saliency map,” which
`topographically codes for local conspicuity over the entire visual
`scene. In primates, such a map is believed to be located in the
`posterior parietal cortex [8] as well as in the various visual maps in
`the pulvinar nuclei of the thalamus [9]. The model’s saliency map
`is endowed with internal dynamics which generate attentional
`shifts. This model consequently represents a complete account of
`
`††††††††††††††††
`•(cid:3) L. Itti and C. Koch are with the Computation and Neural Systems Pro-
`gram, California Institute of Technology—139-74, Pasadena, CA 91125.
`(cid:3)E-mail: {itti, koch}@klab.caltech.edu.
`•(cid:3) E. Niebur is with the Johns Hopkins University, Krieger Mind/Brain Insti-
`tute, Baltimore, MD 21218. E-mail: niebur@jhu.edu.
`Manuscript received 5 Feb. 1997; revised 10 Aug. 1998. Recommended for accep-
`tance by D. Geiger.
`For information on obtaining reprints of this article, please send e-mail to:
`tpami@computer.org, and reference IEEECS Log Number 107349.
`
`0162-8828/98/$10.00 © 1998 IEEE
`
`Fig. 1. General architecture of the model.
`
`bottom-up saliency and does not require any top-down guidance
`to shift attention. This framework provides a massively parallel
`method for the fast selection of a small number of interesting im-
`age locations to be analyzed by more complex and time-
`consuming object-recognition processes. Extending this approach
`in “guided-search,” feedback from higher cortical areas (e.g.,
`knowledge about targets to be found) was used to weight the im-
`portance of different features [10], such that only those with high
`weights could reach higher processing levels.
`
`2 MODEL
`Input is provided in the form of static color images, usually digit-
`ized at 640 · 480 resolution. Nine spatial scales are created using
`dyadic Gaussian pyramids [11], which progressively low-pass
`filter and subsample the input image, yielding horizontal and ver-
`tical image-reduction factors ranging from 1:1 (scale zero) to 1:256
`(scale eight) in eight octaves.
`Each feature is computed by a set of linear “center-surround”
`operations akin to visual receptive fields (Fig. 1): Typical visual
`neurons are most sensitive in a small region of the visual space
`(the center), while stimuli presented in a broader, weaker antago-
`nistic region concentric with the center (the surround) inhibit the
`neuronal response. Such an architecture, sensitive to local spatial
`discontinuities, is particularly well-suited to detecting locations
`which stand out from their surround and is a general computa-
`tional principle in the retina, lateral geniculate nucleus, and pri-
`mary visual cortex [12]. Center-surround is implemented in the
`model as the difference between fine and coarse scales: The center
`is a pixel at scale c ˛ {2, 3, 4}, and the surround is the corresponding
`pixel at scale s = c + d, with d ˛ {3, 4}. The across-scale difference
`between two maps, denoted “*” below, is obtained by interpolation
`to the finer scale and point-by-point subtraction. Using several scales
`not only for c but also for d = s - c yields truly multiscale feature
`extraction, by including different size ratios between the center and
`surround regions (contrary to previously used fixed ratios [5]).
`
`2.1 Extraction of Early Visual Features
`With r, g, and b being the red, green, and blue channels of the in-
`put image, an intensity image I is obtained as I = (r + g + b)/3. I is
`
`Page 1 of 6
`
`BOSCH EXHIBIT 1006
`
`
`
`IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 20, NO. 11, NOVEMBER 1998
`
`1255
`
`Fig. 2. The normalization operator 1(.).
`
`1)(cid:3) normalizing the values in the map to a fixed range [0..M], in
`order to eliminate modality-dependent amplitude differences;
`2)(cid:3) finding the location of the map’s global maximum M and
`computing the average m of all its other local maxima; and
`1
`62 .
`3)(cid:3) globally multiplying the map by M m-
`Only local maxima of activity are considered, such that 1(.)
`compares responses associated with meaningful “activitation
`spots” in the map and ignores homogeneous areas. Comparing the
`maximum activity in the entire map to the average overall activa-
`tion measures how different the most active location is from the
`average. When this difference is large, the most active location
`stands out, and the map is strongly promoted. When the difference
`is small, the map contains nothing unique and is suppressed. The
`biological motivation behind the design of 1(.) is that it coarsely
`replicates cortical lateral inhibition mechanisms, in which neigh-
`boring similar features inhibit each other via specific, anatomically
`defined connections [15].
`Feature maps are combined into three “conspicuity maps,” ,
`for intensity (5), & for color (6), and 2 for orientation (7), at the
`scale (s = 4) of the saliency map. They are obtained through
`across-scale addition, “¯,” which consists of reduction of each
`map to scale four and point-by-point addition:
`2
`7
`3
`8 (5)
`=
`¯ ¯
`1 ,
`
`,
`
`=
`
`4
`
`c
`
`4
`
`=
`
`c
`
`2
`
`s
`
`= +
`c
`
`3
`
`,
`c s
`
`&
`
`=
`
`4
`
`¯ ¯
`
`+
`
`c
`
`4
`
`3
`1 5 *
`
`2
`
`7
`
`8
`
`+
`
`3
`1 % <
`
`2
`
`,
`c s
`
`7
`
`8
`
`. (6)
`
`,
`c s
`
`c
`
`2
`
`s
`
`3
`
`=
`= +
`c
`For orientation, four intermediary maps are first created by
`combination of the six feature maps for a given q and are then
`combined into a single orientation conspicuity map:
`2
`3
`+
`4
`4
`c
`1 2
`
`used to create a Gaussian pyramid I(s), where s ˛ [0..8] is the
`scale. The r, g, and b channels are normalized by I in order to de-
`couple hue from intensity. However, because hue variations are
`not perceivable at very low luminance (and hence are not salient),
`normalization is only applied at the locations where I is larger than
`1/10 of its maximum over the entire image (other locations yield
`zero r, g, and b). Four broadly-tuned color channels are created:
`R = r - (g + b)/2 for red, G = g - (r + b)/2 for green, B = b - (r + g)/2
`for blue, and Y = (r + g)/2 - |r - g|/2 - b for yellow (negative
`values are set to zero). Four Gaussian pyramids R(s), G(s), B(s),
`and Y(s) are created from these color channels.
`Center-surround differences (* defined previously) between a
`“center” fine scale c and a “surround” coarser scale s yield the
`feature maps. The first set of feature maps is concerned with inten-
`sity contrast, which, in mammals, is detected by neurons sensitive
`either to dark centers on bright surrounds or to bright centers on
`dark surrounds [12]. Here, both types of sensitivities are simulta-
`neously computed (using a rectification) in a set of six maps ,(c, s),
`with c ˛ {2, 3, 4} and s = c + d, d ˛ {3, 4}:
`,(c, s) = |I(c) * I(s)|. (1)
`A second set of maps is similarly constructed for the color
`channels, which, in cortex, are represented using a so-called “color
`double-opponent” system: In the center of their receptive fields,
`neurons are excited by one color (e.g., red) and inhibited by an-
`other (e.g., green), while the converse is true in the surround. Such
`spatial and chromatic opponency exists for the red/green,
`green/red, blue/yellow, and yellow/blue color pairs in human
`primary visual cortex [13]. Accordingly, maps 5*(c, s) are created
`in the model to simultaneously account for red/green and
`green/red double opponency (2) and %<(c, s) for blue/yellow and
`yellow/blue double opponency (3):
`5*(c, s) = |(R(c) - G(c)) * (G(s) - R(s))| (2)
`%<(c, s) = |(B(c) - Y(c)) * (Y(s) - B(s))|. (3)
`Local orientation information is obtained from I using oriented
`Gabor pyramids O(s, q), where s ˛ [0..8] represents the scale and
`q ˛ {0o, 45o, 90o, 135o} is the preferred orientation [11]. (Gabor fil-
`ters, which are the product of a cosine grating and a 2D Gaussian
`envelope, approximate the receptive field sensitivity profile (impulse
`response) of orientation-selective neurons in primary visual cortex
`[12].) Orientation feature maps, 2(c, s, q), encode, as a group, local
`orientation contrast between the center and surround scales:
`2(c, s, q) = |O(c, q) * O(s, q)|. (4)
`In total, 42 feature maps are computed: six for intensity, 12 for
`color, and 24 for orientation.
`
`2.2 The Saliency Map
`The purpose of the saliency map is to represent the conspicuity—
`or “saliency”—at every location in the visual field by a scalar quan-
`tity and to guide the selection of attended locations, based on the
`spatial distribution of saliency. A combination of the feature maps
`provides bottom-up input to the saliency map, modeled as a dy-
`namical neural network.
`One difficulty in combining different feature maps is that they
`represent a priori not comparable modalities, with different dy-
`namic ranges and extraction mechanisms. Also, because all 42
`feature maps are combined, salient objects appearing strongly in
`only a few maps may be masked by noise or by less-salient objects
`present in a larger number of maps.
`In the absence of top-down supervision, we propose a map
`normalization operator, 1(.), which globally promotes maps in
`which a small number of strong peaks of activity (conspicuous loca-
`tions) is present, while globally suppressing maps which contain
`numerous comparable peak responses. 1(.) consists of (Fig. 2):
`
`. (7)
`
`(cid:28)(cid:30)(cid:29)
`
`7
`8
`
`q
`
`,
`c s
`
`,
`
`(cid:25)(cid:27)(cid:26)
`
`c
`
`1
`
`2
`
`=
`
`˛ (cid:176) (cid:176) (cid:176) (cid:176) ¯ ¯(cid:229)
`@
`;
`=
`= +
`q
`2
`3
`s c
`,
`,
`,
`0 45 90 135
`The motivation for the creation of three separate channels, , ,
`& , and 2 , and their individual normalization is the hypothesis
`that similar features compete strongly for saliency, while different
`modalities contribute independently to the saliency map. The three
`conspicuity maps are normalized and summed into the final input
`6 to the saliency map:
`=
`6
`
`8
`3
`8
`3
`8
`3
`4
`9. (8)
`
`
`1 , 1 & 1 2+ +
`At any given time, the maximum of the saliency map (SM) de-
`fines the most salient image location, to which the focus of atten-
`tion (FOA) should be directed. We could now simply select the
`most active location as defining the point where the model should
`next attend. However, in a neuronally plausible implementation,
`we model the SM as a 2D layer of leaky integrate-and-fire neurons
`at scale four. These model neurons consist of a single capacitance
`which integrates the charge delivered by synaptic input, of a leak-
`age conductance, and of a voltage threshold. When the threshold is
`
`1 3
`
`Page 2 of 6
`
`
`
`1256
`
`IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 20, NO. 11, NOVEMBER 1998
`
`scale s = 4, in which synaptic interactions among units ensure that
`only the most active location remains, while all other locations are
`suppressed.
`The neurons in the SM receive excitatory inputs from 6 and are
`all independent. The potential of SM neurons at more salient loca-
`tions hence increases faster (these neurons are used as pure inte-
`grators and do not fire). Each SM neuron excites its corresponding
`WTA neuron. All WTA neurons also evolve independently of each
`other, until one (the “winner”) first reaches threshold and fires.
`This triggers three simultaneous mechanisms (Fig. 3):
`1)(cid:3) The FOA is shifted to the location of the winner neuron;
`2)(cid:3) the global inhibition of the WTA is triggered and completely
`inhibits (resets) all WTA neurons;
`3)(cid:3) local inhibition is transiently activated in the SM, in an area
`with the size and new location of the FOA; this not only
`yields dynamical shifts of the FOA, by allowing the next
`most salient location to subsequently become the winner,
`but it also prevents the FOA from immediately returning to
`a previously-attended location.
`Such an “inhibition of return” has been demonstrated in human
`visual psychophysics [16]. In order to slightly bias the model to
`subsequently jump to salient locations spatially close to the cur-
`rently-attended location, a small excitation is transiently activated
`in the SM, in a near surround of the FOA (“proximity preference”
`rule of Koch and Ullman [4]).
`Since we do not model any top-down attentional compo-
`nent, the FOA is a simple disk whose radius is fixed to one-
`sixth of the smaller of the input image width or height. The
`time constants, conductances, and firing thresholds of the
`simulated neurons were chosen (see [17] for details) so that the
`FOA jumps from one salient location to the next in approxi-
`mately 30–70 ms (simulated time), and that an attended area is
`inhibited for approximately 500–900 ms (Fig. 3), as has been
`observed psychophysically [16]. The difference in the relative
`magnitude of these delays proved sufficient to ensure thorough
`scanning of the image and prevented cycling through only a
`limited number of locations. All parameters are fixed in our
`implementation [17], and the system proved stable over time
`for all images studied.
`
`2.3 Comparison With Spatial Frequency Content Models
`Reinagel and Zador [18] recently used an eye-tracking device to
`analyze the local spatial frequency distributions along eye scan
`paths generated by humans while free-viewing gray-scale images.
`They found the spatial frequency content at the fixated locations to
`be significantly higher than, on average, at random locations. Al-
`though eye trajectories can differ from attentional trajectories un-
`der volitional control [1], visual attention is often thought as a pre-
`occulomotor mechanism, strongly influencing free-viewing. It was,
`hence, interesting to investigate whether our model would repro-
`duce the findings of Reinagel and Zador.
`We constructed a simple measure of spatial frequency content
`(SFC): At a given image location, a 16 · 16 image patch is extracted
`from each I(2), R(2), G(2), B(2), and Y(2) map, and 2D Fast Fourier
`Transforms (FFTs) are applied to the patches. For each patch, a
`threshold is applied to compute the number of nonnegligible FFT
`coefficients; the threshold corresponds to the FFT amplitude of a
`just-perceivable grating (1 percent contrast). The SFC measure is
`the average of the numbers of nonnegligible coefficients in the five
`corresponding patches. The size and scale of the patches were cho-
`sen such that the SFC measure is sensitive to approximately the
`same frequency and resolution ranges as our model; also, our SFC
`measure is computed in the RGB channels as well as in intensity,
`like the model. Using this measure, an SFC map is created at scale
`four and is compared to the saliency map (Fig. 4).
`
`Fig. 3. Example of operation of the model with a natural image. Parallel
`feature extraction yields the three conspicuity maps for color contrasts
`(& ), intensity contrasts (, ), and orientation contrasts (2 ). These are
`combined to form input 6 to the saliency map (SM). The most salient
`location is the orange telephone box, which appeared very strongly in
`& ; it becomes the first attended location (92 ms simulated time). After
`the inhibition-of-return feedback inhibits this location in the saliency
`map, the next most salient locations are successively selected.
`
`reached, a prototypical spike is generated, and the capacitive
`charge is shunted to zero [14]. The SM feeds into a biologically-
`plausible 2D “winner-take-all” (WTA) neural network [4], [1] at
`
`Page 3 of 6
`
`
`
`IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 20, NO. 11, NOVEMBER 1998
`
`1257
`
` (1) (2) (3)
`
`Fig. 4. (a) Examples of color images. (b) The corresponding saliency map inputs. (c) Spatial frequency content (SFC) maps. (d) Locations at
`which input to the saliency map was higher than 98 percent of its maximum (yellow circles) and image patches for which the SFC was higher than
`98 percent of its maximum (red squares). The saliency maps are very robust to noise, while SFC is not.
`
`3 RESULTS AND DISCUSSION
`Although the concept of a saliency map has been widely used in
`FOA models [1], [3], [7], little detail is usually provided about its
`construction and dynamics. Here, we examine how the feed-
`forward feature-extraction stages, the map combination strategy,
`and the temporal properties of the saliency map all contribute to
`the overall system performance.
`
`3.1 General Performance
`The model was extensively tested with artificial images to ensure
`proper functioning. For example, several objects of the same shape
`but varying contrast with the background were attended to in the
`order of decreasing contrast. The model proved very robust to the
`addition of noise to such images (Fig. 5), particularly if the prop-
`erties of the noise (e.g., its color) were not directly conflicting with
`the main feature of the target.
`
`Page 4 of 6
`
`
`
`1258
`
`IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 20, NO. 11, NOVEMBER 1998
`
`Fig. 5. Influence of noise on detection performance, illustrated with a 768 · 512 scene in which a target (two people) is salient by its unique color
`contrast. The mean ± S.E. of false detections before target found is shown as a function of noise density for 50 instantiations of the noise. The
`system is very robust to noise which does not directly interfere with the main feature of the target (left; intensity noise and color target). When the
`noise has similar properties to the target, it impairs the target’s saliency and the system first attends to objects salient for other features (here,
`coarse-scale variations of intensity).
`
`The model was able to reproduce human performance for a
`number of pop-out tasks [7], using images of the type shown in
`Fig. 2. When a target differed from an array of surrounding dis-
`tractors by its unique orientation (as in Fig. 2), color, intensity, or
`size, it was always the first attended location, irrespective of the
`number of distractors. Contrarily, when the target differed from the
`distractors only by a conjunction of features (e.g., it was the only red
`horizontal bar in a mixed array of red vertical and green horizontal
`bars), the search time necessary to find the target increased linearly
`with the number of distractors. Both results have been widely ob-
`served in humans [7] and are discussed in Section 3.2.
`We also tested the model with real images, ranging from natu-
`ral outdoor scenes to artistic paintings and using 1(.) to normalize
`
`the feature maps (Fig. 3 and [17]). With many such images, it is
`difficult to objectively evaluate the model, because no objective
`reference is available for comparison, and observers may disagree
`on which locations are the most salient. However, in all images
`studied, most of the attended locations were objects of interest,
`such as faces, flags, persons, buildings, or vehicles.
`Model predictions were compared to the measure of local SFC,
`in an experiment similar to that of Reinagel and Zador [18], using
`natural scenes with salient traffic signs (90 images), a red soda can
`(104 images), or a vehicle’s emergency triangle symbol (64 images).
`Similar to Reinagel and Zador’s findings, the SFC at attended lo-
`cations was significantly higher than the average SFC, by a factor
`decreasing from 2.5 ± 0.05 at the first attended location to 1.6 ± 0.05
`
`Page 5 of 6
`
`
`
`IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 20, NO. 11, NOVEMBER 1998
`
`1259
`
`at the eighth attended location. Although this result does not neces-
`sarily indicate similarity between human eye fixations and the
`model’s attentional trajectories, it indicates that the model, like hu-
`mans, is attracted to “informative” image locations, according to the
`common assumption that regions with richer spectral content are
`more informative. The SFC map was similar to the saliency map for
`most images (e.g., Fig. 4.1). However, both maps differed substan-
`tially for images with strong, extended variations of illumination or
`color (e.g., due to speckle noise): While such areas exhibited uni-
`formly high SFC, they had low saliency because of their uniformity
`(Fig. 4.2 and Fig. 4.3). In such images, the saliency map was usually
`in better agreement with our subjective perception of saliency.
`Quantitatively, for the 258 images studied here, the SFC at attended
`locations was significantly lower than the maximum SFC, by a factor
`decreasing from 0.90 ± 0.02 at the first attended location to 0.55 ±
`0.05 at the eighth attended location: While the model was attending
`to locations with high SFC, these were not necessarily the locations
`with highest SFC. It consequently seems that saliency is more than
`just a measure of local SFC. The model, which implements within-
`feature spatial competition, captured subjective saliency better than
`the purely local SFC measure.
`
`3.2 Strengths and Limitations
`We have proposed a model whose architecture and components
`mimic the properties of primate early vision. Despite its simple
`architecture and feed-forward feature-extraction mechanisms, the
`model is capable of strong performance with complex natural
`scenes. For example, it quickly detected salient traffic signs of var-
`ied shapes (round, triangular, square, rectangular), colors (red,
`blue, white, orange, black), and textures (letter markings, arrows,
`stripes, circles), although it had not been designed for this pur-
`pose. Such strong performance reinforces the idea that a unique
`saliency map, receiving input from early visual processes, could
`effectively guide bottom-up attention in primates [4], [10], [5], [8].
`From a computational viewpoint, the major strength of this ap-
`proach lies in the massively parallel implementation, not only of
`the computationally expensive early feature extraction stages, but
`also of the attention-focusing system. More than previous models
`based extensively on relaxation techniques [5], our architecture
`could easily allow for real-time operation on dedicated hardware.
`The type of performance which can be expected from this model
`critically depends on one factor: Only object features explicitly rep-
`resented in at least one of the feature maps can lead to pop-out, that
`is, rapid detection independent of the number of distracting objects
`[7]. Without modifying the preattentive feature-extraction stages,
`our model cannot detect conjunctions of features. While our system
`immediately detects a target which differs from surrounding dis-
`tractors by its unique size, intensity, color, or orientation (properties
`which we have implemented because they have been very well
`characterized in primary visual cortex), it will fail at detecting tar-
`gets salient for unimplemented feature types (e.g., T junctions or line
`terminators, for which the existence of specific neural detectors re-
`mains controversial). For simplicity, we also have not implemented
`any recurrent mechanism within the feature maps and, hence,
`cannot reproduce phenomena like contour completion and closure,
`which are important for certain types of human pop-out [19]. In addi-
`tion, at present, our model does not include any magnocellular motion
`channel, which is known to play a strong role in human saliency [5].
`A critical model component is the normalization 1(.), which
`provided a general mechanism for computing saliency in any situa-
`tion. The resulting saliency measure implemented by the model,
`although often related to local SFC, was closer to human saliency,
`because it implemented spatial competition between salient locations.
`Our feed-forward implementation of 1(.) is faster and simpler than
`previously-proposed iterative schemes [5]. Neuronally, spatial com-
`petition effects similar to 1(.) have been observed in the nonclassi-
`
`cal receptive field of cells in striate and extrastriate cortex [15].
`In conclusion, we have presented a conceptually simple com-
`putational model for saliency-driven focal visual attention. The
`biological insight guiding its architecture proved efficient in re-
`producing some of the performances of primate visual systems.
`The efficiency of this approach for target detection critically de-
`pends on the feature types implemented. The framework pre-
`sented here can consequently be easily tailored to arbitrary tasks
`through the implementation of dedicated feature maps.
`
`ACKNOWLEDGMENTS
`We thank Werner Ritter and Daimler-Benz for the traffic sign images
`and Pietro Perona and both reviewers for excellent suggestions.
`This research was supported by the U.S. National Science
`Foundation, the Center for Neuromorphic Engineering at Caltech,
`and the U.S. Office of Naval Research.
`
`[8](cid:3)
`
`REFERENCES
`[1](cid:3)
`J.K. Tsotsos, S.M. Culhane, W.Y.K. Wai, Y.H. Lai, N. Davis, and F.
`Nuflo, “Modelling Visual Attention via Selective Tuning,” Artifi-
`cial Intelligence, vol. 78, no. 1-2, pp. 507–545, Oct. 1995.
`[2](cid:3) E. Niebur and C. Koch, “Computational Architectures for Atten-
`tion,” R. Parasuraman, ed., The Attentive Brain, pp. 163–186. Cam-
`bridge, Mass.: MIT Press, 1998.
`[3](cid:3) B.A. Olshausen, C.H. Anderson, and D.C. Van Essen, “A Neuro-
`biological Model of Visual Attention and Invariant Pattern Rec-
`ognition Based on Dynamic Routing of Information,” J. Neurosci-
`ence, vol. 13, no. 11, pp. 4,700–4,719, Nov. 1993.
`[4](cid:3) C. Koch and S. Ullman, “Shifts in Selective Visual Attention: To-
`wards the Underlying Neural Circuitry,” Human Neurobiology,
`vol. 4, pp. 219–227, 1985.
`[5](cid:3) R. Milanese, S. Gil, and T. Pun, “Attentive Mechanisms for Dy-
`namic and Static Scene Analysis,” Optical Eng., vol. 34, no. 8,
`pp. 2,428–2,434, Aug. 1995.
`[6](cid:3) S. Baluja and D.A. Pomerleau, “Expectation-Based Selective Atten-
`tion for Visual Monitoring and Control of a Robot Vehicle,” Robotics
`and Autonomous Systems, vol. 22, no. 3-4, pp. 329–344, Dec. 1997.
`[7](cid:3) A.M. Treisman and G. Gelade, “A Feature-Integration Theory of
`Attention,” Cognitive Psychology, vol. 12, no. 1, pp. 97–136, Jan. 1980.
`J.P. Gottlieb, M. Kusunoki, and M.E. Goldberg, “The Representa-
`tion of Visual Salience in Monkey Parietal Cortex,” Nature, vol. 391,
`no. 6,666, pp. 481-484, Jan. 1998.
`[9](cid:3) D.L. Robinson and S.E. Peterson, “The Pulvinar and Visual Sali-
`ence,” Trends in Neurosciences, vol. 15, no. 4, pp. 127–132, Apr. 1992.
`[10](cid:3) J.M. Wolfe, “Guided Search 2.0: A Revised Model of Visual
`Search,” Psychonomic Bull. Rev., vol. 1, pp. 202–238, 1994.
`[11](cid:3) H. Greenspan, S. Belongie, R. Goodman, P. Perona, S. Rakshit, and
`C.H. Anderson, “Overcomplete Steerable Pyramid Filters and
`Rotation Invariance,” Proc. IEEE Computer Vision and Pattern Rec-
`ognition, pp. 222-228, Seattle, Wash., June 1994.
`[12](cid:3) A.G. Leventhal, The Neural Basis of Visual Function: Vision and Vis-
`ual Dysfunction, vol. 4. Boca Raton, Fla.: CRC Press, 1991.
`[13](cid:3) S. Engel, X. Zhang, and B. Wandell, “Colour Tuning in Human
`Visual Cortex Measured With Functional Magnetic Resonance
`Imaging,” Nature, vol. 388, no. 6,637, pp. 68–71, July 1997.
`[14](cid:3) C. Koch, Biophysics of Computation: Information Processing in Single
`Neurons. New York: Oxford Univ. Press, 1998.
`[15](cid:3) M.W. Cannon and S.C. Fullenkamp, “A Model for Inhibitory Lat-
`eral Interaction Effects in Perceived Contrast,” Vision Res., vol. 36,
`no. 8, pp. 1,115–1,125, Apr. 1996.
`[16](cid:3) M.I. Posner and Y. Cohen, “Components of Visual Orienting,”
`H. Bouma and D.G. Bouwhuis, eds., Attention and Performance,
`vol. 10, pp. 531–556. Hilldale, N.J.: Erlbaum, 1984.
`[17](cid:3) The C++ implementation of the model and numerous examples of
`attentional predictions on natural and synthetic images can be
`retrieved from http://www.klab.caltech.edu/~itti/attention/.
`[18](cid:3) P. Reinagel and A.M. Zador, “The Effect of Gaze on Natural Scene
`Statistics,” Neural Information and Coding Workshop, Snowbird,
`Utah, 16-20 Mar. 1997.
`[19](cid:3) I. Kovacs and B. Julesz, “A Closed Curve Is Much More Than an
`Incomplete One: Effect of Closure in Figure-Ground Segmenta-
`tion,” Proc. Nat’l Academy of Sciences, U.S.A., vol. 90, no. 16, pp. 7,495–
`7,497, Aug. 1993.
`
`Page 6 of 6
`
`