Computational Cognitive Neuroscience Approach for Saliency ... · Computational Cognitive Neuroscience Approach for Saliency Map detection using Graph-Based Visual Saliency (GBVS)

Computational Cognitive NeuroscienceApproach for Saliency Map detectionusing Graph-Based Visual Saliency(GBVS) tool in Machine Vision

Manjunath R Kounte11 and B K Sujatha2

1School of Electronics and Communication EngineeringREVA University, Bangalore, Karnataka, India

[email protected] Engineering,

M S Ramaiah Institute of Technology,Bangalore, [email protected]

January 1, 2018

Abstract

Objective: In this paper, we propose computational cog-nitive neuroscience based visual saliency model called asGraph based Visual Saliency (GBVS). The Objectives in-cludes designing and implementing bottom up model of vi-sual attention, wherein it tries to predict the human eyefixation points by taking real time digital images chosenrandomly from everyday life as input.

Methods / Statistical Analysis: The methodology usedhere supports biological modelling and parallelized natu-rally, the process involves identifying feature channels, thenforming of activation maps and finally normalizing all threein a way which highlights similarity of conspicuity and com-bination with other maps. The basic evaluation technique

1

International Journal of Pure and Applied MathematicsVolume 118 No. 16 2018, 1207-1226ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)url: http://www.ijpam.euSpecial Issue ijpam.eu

1207

used here is the calculation of the ROC using graph basedvisual saliency compared with the other techniques.

Findings: In the implementation, after taking digital in-put as input, feature maps are evaluated, followed by cal-culation of the saliency map. In the results, the foregroundand background is achieved by fixing the threshold valuesfor the saliency map points. The main parameter used forperformance evaluation of analyzing human eye fixation wasROC. The ROC measurement is better than the ROC valueevaluated for the traditional saliency map computations.The ROC value for GBVS technique is 0.74, whereas theRIC value for traditional saliency map is 0.57.

Application / Improvement: Applications for visual at-tention, reinforcement learning, adaptation, sensory pro-cessing involves the basic concept of surprise at its core.We present their unique way of mathematical analysis forsurprise theory using the Bayesian modelling. The surprisegenerated by inducement or an event needs to be character-ized quantitatively by using mathematical modelling for alldata coming from multifaceted natural environments.

Key Words : Visual Attention, Graph Based VisualSaliency (GBVS), Saliency Map, Computational CognitiveNeuroscience (CCNS), Visual Attention.

1 Introduction

Machine vision provides us a new area of research and in generalvision community research includes an area where predicting thehuman eye fixating at a particular scenario. The human beingsuse this ability to view a scene to identify in detail the most im-portant and relevant areas, while the amount and energy spent onprocessing of these information is very less.

William James said Everyone knows what attention is. It isthe taking possession by the mind, in clear and vivid form, of oneout of what seem several simultaneously possible objects or trainsof thought. Focalization, concentration, of consciousness are of itsessence. It implies withdrawal from some things in order to dealeffectively with others1.

One of the most severe problems of perception is information

2

International Journal of Pure and Applied Mathematics Special Issue

1208

overload. Peripheral sensors generate afferent signals more or lesscontinuously and it would be computationally costly to process allthis incoming information all the time 2−3. Thus, it is important forthe nervous system to make decisions on which part of the availableinformation is to be selected for further, more detailed processing,and which parts are to be discarded. Furthermore, the selectedstimuli need to be prioritized, with the most relevant being pro-cessed first and the less important ones later, thus leading to asequential treatment of different parts of the visual scene. This se-lection and ordering process is called selective attention. Amongmany other functions, attention to a stimulus has been considerednecessary for it to be perceived consciously4−5.

What determines which stimuli are selected by the attentionalprocess and which will be discarded? Many interacting factors con-tribute to this decision. The decision of distinguishing between se-lections and discarding the information based on the stimuli helpsin classifying top-down as well as bottom-up factors. The bottom-up control factors are mainly dependent on the instantaneous sen-sory input, without taking into account the internal state of theorganism6. Top-down control, on the other hand, does take intoaccount the internal state, such as goals the organisms has at thistime, personal history and experiences, etc7. A dramatic exampleof a stimulus that attracts attention using bottom-up mechanismsis a fire-cracker going off suddenly while an example of top-down at-tention is the focusing onto difficult-to-find food items by an animalthat is hungry, ignoring more ”salient” stimuli8−9.

One of the most important mechanism which attention pro-vides is the neglecting of interference from parallel events duringthe process of selection of relevant features of a scene for subse-quent processing. A common misinterpretation is that attentionand ocular fixation are one and the same phenomenon. Atten-tion focuses processing on a selected region of the visual field thatneedn’t coincide with the center of fixation10−11.Visual attention isan appliance of the human visual system to highlight on certainportions of a scene first, before attention is fatigued on to the otherparts. Such areas that capture primary attention are called visualattention regions (VARs). For various multimedia applications, theVisual Attention Region (VAR) should then indeed be the regionsof interest and an automatic practice of extracting the VARs be-

3


1209

comes necessary12−13. Documentation of VARs has been presentedto be useful for object recognition and region based image retrieval.Similarly, images can be adapted for different users with differentdevice proficiencies based on the VARs extracted from the image,thus improving viewing choice. Examples of such version compriseinvoluntary surfing for large images, image resolution adaptationand automatic thumbnail generation14.

Feature Selection has received increasing attention in machinelearning research. Even though the ever upgrading memory andCPU configurations have made nowadays computers more and morepowerful, the amount of information, and sometimes the large di-mensionalities of datasets, still challenge the efficiency and effec-tiveness of conventional algorithms.

Section II gives information on extracting early visual featuresand visual processing from the retina. Section III describes SaliencyMap detection and hypothesis for extraction of saliency map andidentifying the visual attention region, Section IV discusses GraphBased Visual Saliency design and implementation, followed by dis-cussion on Results, Conclusion.

2 Extracting Early Visual Features

1.1. Visual Processing

Figure 1: Visual Processing from the retina through lateral genic-ulate nucleus of the thalamus to primary visual cortex.

Figure 1, shows the graphical representation of basic optics andtransmission corridors of input visual signals, which entre through

4


1210

the retina, and headway to the lateral geniculate nucleus of thethalamus (LGN), and further to primary visual cortex (V1)15.Theprimary organizing principles at work here, and in other perceptualmodalities and perceptual areas more generally, are: i) Transduc-tion of different information – in the retina, photoreceptors aresensitive to different wavelengths of light (red = long wavelengths,green = medium wavelengths, and blue = short wavelengths), giv-ing us color vision, but the retinal signals also differ in their spatialfrequency (how coarse or fine of a feature they detect – photorecep-tors in the central fovea region can have high spatial frequency =fine resolution, while those in the periphery are lower resolution),and in their temporal response (fast vs. slow responding, includingdifferential sensitivity to motion).

3 Hypothesis for Saliency Map Detec-

tion

1.2. Taxonomy of Saliency Detection MethodsSaliency estimation methods can broadly be classified as into

three categories as shown in fig 2, namely General Computationalmodelling, Hybrid Modelling and Computational Cognitive Neuro-science Modelling.

Figure 2: Taxonomy of Saliency Detection Methods.

The goal of any method, in general, is to detect propertiesof contrast, rarity, or unpredictability in images, of a central re-gion with its surroundings, either locally or globally, using one

5


1211

or more low-level features like color, intensity, and orientation.General computational methods mainly rely on principles of, spec-tral domain processing, information theory or signal processing forsaliency detection16−17. In Hybrid individual feature maps are cre-ated separately and then combined to obtain the final saliency map,while in some others, a combined-feature saliency map is computed.Computational Cognitive Neuroscience model based methods at-tempt to mimic known models of the visual system for detectingsaliency. Some of these algorithms detect saliency over manifoldscales, while others work on a single scale.

1.3. Itti and Koch ModelThe model proposed in 18−19 is shown in Figure.3. The Model is

inspired from the feature integration theory, which explains humanvisual search strategies. First, Visual input is decomposed into aset of topographic feature maps. All feature maps which then in abottom-up manner, combine into a saliency map. This model conse-quently represents a complete interpretation of bottom-up saliencyand does not require any top-down mechanism to change the atten-tion. This framework provides an immensely parallel method forthe fast selection of a small number of attention-grabbing imagelocations to be explored by object-recognition processes20.

Figure 3: Koch and Itti Model.

6


1212

3.3 Bayesian Theory of SurpriseThe surprise in general can be can be explained or defined by

using the following two principles. First, it is said be available onlybecause of uncertainty, because the uncertainty comes from intrin-sic randomness, unavailability of information, or lack of resourcesused for computing. Example: if the real world is mathematicallypredictable and deterministic in nature, then obviously there is nosurprise in it. Second, surprise can never be absolute, its can only beexplained relatively in terms of expectations of the visual observer,or a machine observer. The given input data of visual stimulus canresult varying surprises for different individuals or machines, simi-larly the same input visual input data for an individual at varyingtimes may result in different surprises21. Therefore the Bayesiantheory of surprise is defined based on relativity rather than abso-lute definition, the optimum way to mathematically model wouldbest be in terms and decision theory and probability. Its definitioninvolve (1) to define surprise and randomness in terms of proba-bilistic concepts; (2) Earlier (prior) data and later (posterior) datadistributions in terms of decision theory22.

Fig 4, shows the Computing surprise in early sensory neurons.Fig 4(a) shows the earlier (Prior) data observations, whereas fig4(b) shows the standard Bayesian fitting model23.

As common with this Bayesian approach, the contextual infor-mation of an observer is apprehended by his/her/its earlier (prior)data probability distribution P(H)HS where the H represents hy-pothesis over the space S

The observer D now has to convert this prior data P(H)HS intoposterior data P(H—D) HS via Bayes theorem.

Whereby

∀HεS, {P (H|D)} = P (DVH)P (D)

P (H) (1)

In this context, the new data observation D is said to containno surprise if the earlier data and later data are similar; and it issaid to contain surprise if the earlier data (prior) of the observer Dcontains significantly different data than the later (posterior) data.Therefore the relative of prior and posterior data helps in definingthe surprise, As per Kullback (KL) separation24, surprise is defined

7


1213

as follows:

SUR(D,S) = KL(P (H|D), P (H)) =∫sP (H|D)log P (HVD)

P (H)dH

Figure 4: Computing surprise in early sensory neurons.

4 Graph Based Visual Saliency

4.3.1 Computing Feature MapsThe Fig 4, shows the visual features are derived from the ap-

plied digital nature visual input image and separated into three newdifferent passages with run parallel to each other. After this extrac-tion and an individual management, a feature map is computed foreach channel.

The first feature map, termed as Intensity Channel is obtainedfrom the red color, green color and blue color channels of the inputimage. If Ir, Ig and Ib are assumed to be the red color channel,

8


1214

green color channel and blue color channels respectively, then thetotal intensity It, is given by taking average of all three channels.

It = ( Ir + Ig + Ib ) /3 (3)

The total intensity It is the then used to compute the Gaussianpyramid It(α). Where is the total number of down samples usedin low pass filtering process.

The evaluation of luminance is varies due to encoding values orrepresentations. It is 30% due to red, 59% due to green and 11%due to blue.

The Second feature map, Color Channel is constructed by tak-ing into consideration pyramids as shown Fig 5. The colors red,green and blue are normalized (mathematical Process) by It in or-der to separate hue (shade) from the intensity.

Four color channels are formed as follows:

CR = Ir - (Ig+Ib)/2 (4)is formed for the color red.

CG = Ig - (Ir+Ib)/2 (5)is formed for the color green.

CB = Ib - (Ir+Ig)/2 (6)is formed for the color blue.

CY=Ir+Ig-2(|Ir− Ig|+Ib) (7)is formed for the color yellow.

Gaussian Pyramids are now created for all four colors. They arerepresented as GR(α), GG(α), GB(α) and GY(α). Fig 5, showsthe implementation of the 4 color channels and behaviour of thechromatic activity of these channels.

The third feature map, Local orientation is computed from thetotal intensity It using Gabor pyramids Channels GPO (σ) are cal-culated and evaluated by performing convolution of the altitudes ofthe intensity values evaluated earlier with Gabor filters: GF(α, σ),where αε[0...80%] represents the scale and ωε{0o, 45o, 90o, 135o} is

9


1215

the favored orientation. The feature maps considered due to theOrientation are denoted by, OF(ce,su,ω), encode, as a group, withthe orientation for the nearby locations(local) varies in betweenthe main center(focused) and nearby surrounding values. Later,the Center-surround approachable fields are replicated by a cross-scale numerical subtraction among two maps at the center and theneighboring (surround) levels in these pyramids, thus resulting inthe calculation of the yet another feature map. Later saliency mapis extracted from the feature maps.

Figure 5: Feature Map Fusion.

4.3.2 Forming an Activation MapThere are three feature maps generated which include the for-

mation of intensity, color and orientation. Lets assume the each

10


1216

feature map is represented as F: [n]2 → R. The next most impor-tant objective is to evaluate the activation map A: [n]2 → R, suchthat, naturally, the locations points are denoted by (x,y) ε [n]2where It, or as a substitution, M (x,y); will be represented withrespect to its neighborhood. If the values are high, it representslarger value of the activation map A.

The Markovian Approach, is one of the better way to representthe dissimilarity index, its a more rational way to represent com-pared to any other representations. Now, dissimilarity is definedwith respect to M(x,y) with respect to M(r,s)

d((x, y)||(r, s)) ∆=∣∣∣ logM(x,y)

logM(r,s)

∣∣∣ (8)

The above definition above indicate the dissimilarity index. Itgives the ratio between the two entities. The dissimilarity indexabout the two entities is characterized in terms of logarithmic scale.

The exemplification can also be done in terms of taking thedifference between the M(x,y) and M (r,s). The representation interms of logarithmic scale and taking the difference both works well.

In graph based visual saliency, let us consider the full graphwhich consists of nodes connected with each other. We assumethat its a directed graph and all the nodes are connected to eachother. Let us denote the graph as GVS. Now each node M with itsindices (x,y) ε [n]2 is connected to all other nodes.

The edge point from a node in 2-dimensional plane (x, y) tonode (r, s) will be given as a numerical value i.e. weight as follows:

w1((x, y), (r, s)∆= d((x, y)||(r, s)).F (x−r, y−s) (9)

F (m,n)∆= exp(−(m2+n2)

2σ2) (10)

is a free constraint of the Graph Based Visual Saliency (GBVS)algorithm 25−26.

The next constraint to be discussed here is the weight of the edgefrom node (x,y) to node (r,s) is directly proportional to dissimilarityindex. The markovian approach suggests that the normalization ofthe graph GVS has to be performed by taking the normalizationfactor with each node to 1.

We call the result as cognitive based as the nodes M (in this

11


1217

case neurons) are available in a connected and well organised man-ner, the directed of all nodes is equivalent to the entire network(i.e. visual cortex). The fig 6, shows the information measure andentropy framework.

Figure 6: Framework for Information Measure.

4.3.3 ”Normalizing” an Activation MapThere are numerous means to mix or combine the feature map

details extracted using the center- surround mechanisms process asdiscussed in the previous subsection. One of the many accessibletechniques like non-linear amplification, localized iterations etc isselected and the operator is denoted as N (.). Normalize the nu-merical values in all the feature maps to a particular fixed range[0...GM] so that the modality-dependent generosity differences canbe eliminated. Find the location of the feature maps global maxi-mum GM and compute the average maximum g

′m of all its other

maximums from the local area; and globally multiplying the mapby(GM-g

′m)2.

In the graph based visual saliency (GBVS) algorithm, one of themain objective of the ”normalization” step is certain and a clearstandard process as compared to the previous activation step. Re-cent trends in the process of activation and normalization suggeststhat its a critical area of interest in many aspects. The map re-

12


1218

sulted after the normalization process earlier to the mathematicaladdition combination, may end up with containing less or no infor-mation. Thus, the basic objective of this step would be identify theattention quantity on activation maps27.

Since the attention quantity on activation maps is available, themarkovian analysis will be as follows for the normalization step:The activation map is given by A: [n]2 → R, which needs to be”normalized”. The next step would be to build the graph GN withn2 neurons, i.e. nodes are marked with indexing starting from [n]2.For each of the any node (x, y) and every other node (r,s) (includ-ing (x,y)) to which it is linked) introduce an edge from (x, y) to (r,s) with weight:

w2((x, y), (r, s))∆= A(r, s).F (x−r; y−s) (11)

Finally the iterations are used in the normalization process. Themarkovian approach used provides the advantage of steadiness dis-tribution for all nodes, along with the assumption of attention quan-tity weights of edges for each node as 1. If activation is large, theattention quantity will flow automatically. This algorithm as sameadvantages as that others28.

4.4 Results and DiscussionThe basic evaluation technique used here is the calculation of

the ROC using graph based visual saliency compared with the othertechniques. The objective in to try to predict the human eye fixa-tion points by taking real time digital images chosen randomly fromeveryday life as input. As discussed above in the previous steps:the process used is simple, first taking the input image which isdigital in nature, perform the feature map selection process usingknown techniques, then second step includes identifying the acti-vation maps and last step is normalizing the data by the factor ofone, using iteration technique.

Fig 7, shows the saliency map detection based on graphs. Theiteration technique is used for normalization process, and the al-gorithm is denoted by It as described earlier also. For experimentvalidation, the saliency toolbox is used.

For analyzing the performance, first digital image is taken asinput, feature maps are evaluated, then then saliency maps is calcu-

13


1219

lated, then the target locations are given, i.e. in a real time naturalimage, what normally a human eye focuses on, called human eyefixation. Then, a threshold value is fixed as an evaluation parame-ter, if the saliency map points goes above the threshold value, it isclassified as foreground or target achieved, else if points are belowthe threshold value, it is classified as background. The one moreperformance metric considered for analyzing human eye fixation bythe saliency map is ROC.

The fig 8(a), shows the image for this performance analysis. Theimage samples are collected from the above. Each image is croppedfor the resolution 600x400 pixels, to maintain fair comparison, thefirst step involved using same algorithm to compute the featuremaps. The next involved finding the orientation maps, again tomaintain the same comparison, the angles used were also same ω ={00, 450, 900, 1350}. The markings xs provides the information ofthe human eye fixation points.

Fig 8(b) shows the output when using graph based visual saliency,the ROC measurement is better than the ROC value evaluated forthe traditional saliency map computations. The ROC value forGBVS technique is 0.74, whereas the ROC value for traditionalsaliency map (fig 8(c)) is 0.57. The calculation of center-surroundmaps by finding the difference between the given digital image fea-ture map value and the feature map evaluated after the activationmap.

Figure 7: Saliency Detection based on graphs.

14


1220

Figure 8: a) Sample Picture with Fixation b) Graph-Based SaliencyMap c) Traditional Saliency Map.

15


1221

5 Conclusion

As researchers we all welcome a new novel, easy to approach solu-tion for any existing research problems, however we must also seekto answer the scientific question of how it is possible that, givenaccess to the same feature information, GBVS predicts human fix-ations more reliably than the standard algorithms. The first obser-vation is that, because r point along the image periphery, it is anemergent property that GBVS promotes higher saliency values inthe center of the image plane. In this paper, we have designed andimplemented a visual saliency model using computational cognitiveneuroscience approach called Graph-Based Visual Saliency (GBVS)which provides a new prospective in modelling of visual saliency,the implementation is quite simple in its approach, supports bi-ological modelling and parallelized naturally, the process involvesidentifying feature channels, then forming of activation maps andfinally normalizing all three in a way which highlights similarity ofconspicuity and combination with other maps.

Also, we have presented how computational cognitive neuro-science approach is the best approach in order for the implementa-tion of saliency map technique for identification of visual attentionregions using Graph-Based Visual Saliency (GBVS) tool in machinevision.

References

[1] W. James, The principles of psychology, New York: Holt, 1890,pp. 403-404.

[2] Lee DK, Itti L, Koch C, Braun J. Attention activates winner-take-all competition among visual filters, Nature neuroscience.1999 Apr 1;2(4):375-81.

[3] Einhuser W, Kruse W, Hoffmann KP, Knig P. Differences ofmonkey and human overt attention under natural conditions,Vision research. 2006 Apr 30;46(8):1194-209.

[4] Itti L, Baldi PF. Bayesian surprise attracts human attention.InAdvances in neural information processing systems, 2005,pp. 547-554.

16


1222

[5] Bruce N, Tsotsos J. Saliency based on information maximiza-tion. InAdvances in neural information processing systems,2005, pp. 155-162.

[6] Parkhurst D, Law K, Niebur E. Modeling the role of salience inthe allocation of overt visual attention. Vision research. 2002Jan 31;42(1):107-23.

[7] Ahmed MR, Sujatha BK. A review of reinforcement learn-ing in neuromorphic VLSI chips using computational cognitiveneuroscience, International Journal of Advanced Research inComputer and Communication Engineering (IJARCCE). 2013Aug;2(8):3558-63.

[8] Ahmed MR, Sujatha BK. Learning and Memory issuesin Neuromorphic Engineering: A Review, InternationalJournal of Applied Engineering Research (IJAER). 2015Oct;10(12):23769-23778.

[9] Ahmed MR, Sujatha BK. Levels of Abstraction and Implemen-tation Issues in Neuromorphic Engineering, 2015 Sep;7(4):555-563.

[10] Banu T, Kounte M. Performance Analysis of Irreversible andReversible Counter, HCTL Open Science and Technology Let-ters (STL): Volume 1, June 2013. 2013 Jun 30:69.

[11] Soumyalatha N, Ambhati RK, Kounte MR. Performance eval-uation of ip wireless networks using two way active measure-ment protocol, InAdvances in Computing, Communicationsand Informatics (ICACCI), 2013 International Conference on2013 Aug 22 pp. 1896-1901.

[12] Kounte MR, Sujatha DB. Top-Down Approach for ModellingVisual Attention using Scene Context Features in MachineVision, International Journal of Applied Engineering Re-search.2015 Oct;10(12):31585-31594.

[13] Kounte MR, Sujatha BK. Identification of Visual AttentionRegions in Machine Vision Using Saliency Map, Interna-tional Conference on Communications and Signal Processing(ICCSP), 2015. Apr 2. pp. 0639-0643.

17


1223

[14] Kounte MR. Dr. B. K. Sujatha. Bottom up Approach forModelling Visual Attention Using Saliency Map in Ma-chine vision: A Computational Cognitive Neuroscience Ap-proach, International Journal of Applied Engineering Research.2015;10(11):30153-30166.

[15] OReilly, R. C., Munakata, Y., Frank, M. J., Hazy, T. E., andContributors. Computational Cognitive Neuroscience, WiKiBook, 1st Edition. 2012.

[16] B. Mohd Jabarullah, C. Nelson Kennedy Babu. Segmenta-tion Using Saliencycolor Mapping Technique. Indian Journalof Science and Technology, 2015, 8(15), 1-5.

[17] Shailesh Kishore, Ramneek Singh Grover, N. Sandeep. ObjectSeparation Using Saliency Algorithm. Indian Journal of Sci-ence and Technology, 2015, 8(S2), 231-234.

[18] Itti L, Koch C. Computational modelling of visual attention.Nature reviews neuroscience, 2001 Mar 1;2(3),194-203.

[19] Itti L, Koch C, Niebur E. A model of saliency-based visual at-tention for rapid scene analysis IEEE Transactions on patternanalysis and machine intelligence. 1998 Nov 1;20(11),1254-1259.

[20] Borji A, Itti L. State-of-the-art in visual attention modeling,IEEE transactions on pattern analysis and machine intelli-gence. 2013 Jan;35(1), 185-207.

[21] Itti L, Baldi PF. Bayesian surprise attracts human attention,In Advances in neural information processing systems 2005.pp. 547-554.

[22] Borji A, Itti L. Bayesian optimization explains human activesearch, In Advances in neural information processing systems2013 pp. 55-63.

[23] Zhang L, Tong MH, Marks TK, Shan H, Cottrell GW. SUN: ABayesian framework for saliency using natural statistics, Jour-nal of vision. 2008 May 3;8(7):32-.

18


1224

[24] Kullback, S. Information Theory and Statistics InformationTheory and Statistics,1959.

[25] Harel J, Koch C, Perona P. Graph-based visual saliency. InAdvances in neural information processing systems 2006. pp.545-552.

[26] Rutishauser U, Walther D, Koch C, Perona P. Is bottom-upattention useful for object recognition?. In Computer Visionand Pattern Recognition, 2004. CVPR 2004. Proceedings ofthe 2004 IEEE Computer Society Conference on 2004 Jul 4(Vol. 2, pp. II-37).

[27] Walther D, Rutishauser U, Koch C, Perona P. On the useful-ness of attention for object recognition, In Workshop on Atten-tion and Performance in Computational Vision at ECCV 2004May (pp. 96-103).

[28] Walther D, Rutishauser U, Koch C, Perona P. Selective visualattention enables learning and recognition of multiple objects incluttered scenes, Computer Vision and Image Understanding.2005 Nov 30;100(1):41-63.

19


1225

1226

Computational Cognitive Neuroscience Approach for Saliency ... · Computational Cognitive Neuroscience Approach for Saliency Map detection using Graph-Based Visual Saliency (GBVS)

Documents