CN101751671B - Visual attention computing method and system based on bit entropy rate - Google Patents

Abstract

The invention discloses a visual attention computing method and a system based on bit entropy rate. The method includes that: a sparse codebook is learned to obtain a sparse codebook primary function; the sparse codebook primary function is adopted to filter the image or video data and obtain a plurality of sub-band feature graphs; a corresponding fully-connected graph is established for each sub-band feature graph; the information transfer method of each fully-connected graph adopts the random walk method, during the random walk process, the significance measurement is performed according to the bit entropy rate to accordingly obtain a bit entropy rate graph corresponding to each sub-band feature graph; the bit entropy rate graphs are added together to obtain the significance graph of the image or video data. As indicated in large amount of experiments, and compared with other methods in prior art, the analysis result of image significance analysis or video significance analysis obtained through the invention is more accurate and is supported by the basis of physiology and psychology.

Description

Visual attention computing method and system based on bit entropy rate

Technical field

The present invention relates to image and video processing technique, a plurality of fields such as computer vision and visually-perceptible relate in particular to a kind of visual attention computing method and system based on bit entropy rate.

Background technology

Selective attention is meant that psychological resource selectively is assigned to some cognitive course, make these cognitive processes to information Processing more quick and precisely.Attention is extremely important for coordinating various cognitive course.Human each moment all receives a large amount of external informations, is in by in the state of " INFORMATION BOMB ".Psychological resource that we are limited and neural resource can not be handled much more so information simultaneously, can only optionally handle the information with high priority and ignore the information of low priority, and the effect of attention is embodied in this just.

Selective attention is a very complicated cognitive process, and it at every moment affects the process of information processing of brain.Research to selective attention has been the hot fields of cognitive science since the eighties of last century the eighties always, and this point can obtain proof from the quality and quantity about the article noted that deliver every year.Theoretical numerous and complicated about attention mechanism, but the researcher relatively admits following classification: divide from noting the direction that produces, selective attention comprises from bottom to top, the process of data-driven (bottom-up and data-driven process) and from top to bottom, the process (top-down and goal-directed process) that target is guided; Divide from the target of note selecting, selective attention comprises the attention (space-basedattention) based on the space, based on the attention (feature-based attention) of feature with based on the attention (object-based attention) of object.

Selective attention generally comprises static noted analyzing and dynamically notes analyzing, respectively at the detection of marking area in still image and the dynamic video.In static state is noted analyzing two kinds of models are arranged, a kind of is the irrelevant model of task from bottom to top, and being also referred to as stimulates the model that drives; Another kind is the relevant model of task from top to bottom, is also referred to as the model of task-driven.Method from bottom to top mainly comes from some achievements in research of visual cognition psychology aspect.Koch and Ullman have just proposed a method simply from bottom to top in 1985, as input and be encoded into significantly figure of an explicit two dimension, the order that conspicuousness descends in this figure provides a kind of effective attention scan mode the stimulation conspicuousness of each position in visual scene response for it.Method from bottom to top need not to consider any knowledge information, yet we are constantly in the influence that is subjected to its knowledge, culture background.In other words, we at every moment are task-driven, observe surroundings with top-down method.Therefore, existing a few thing is applied to priori to improve the degree of agreement of analysis result and eye movement experimental result in the vision attention analysis.In recent years, increasing work begins to be devoted to dynamically to note analyzing.In video, the significance of each frame distributes and not only is subjected to himself content influence, is subjected to the influence of time context relation simultaneously, so motion feature plays crucial effects in video.

There is the problem of two aspects in existing method: on the one hand, existing method has plenty of and has plenty of based on the information maximization principle based on the model of center-on every side, seldom has method to explain from this two aspect simultaneously; On the other hand, existing method is not accurate especially to the prediction of vision attention point, and there is a big difference with real eye movement data.

Summary of the invention

The objective of the invention is to, a kind of visual attention computing method and system are provided,, can obtain the significance analysis result of accurate more image or video, meet physiology and psychologic foundation more based on the present invention.

The invention provides a kind of visual attention computing method based on bit entropy rate, comprise the steps: filter step, adopt sparse code book basis function that image or video data are carried out filtering, obtain a plurality of subband feature figure, wherein said sparse code book basis function is based on the sparse code book of study and obtains; Full connection layout establishment step, each the subband feature figure that is respectively described a plurality of subband feature figure sets up corresponding full connection layout; Bit entropy rate figure obtaining step adopts the method for random walk to carry out the transmission of information on each described full connection layout, in the random walk process, carries out significance tolerance according to bit entropy rate, and then obtains the bit entropy rate figure of each described subband feature figure correspondence; The saliency map obtaining step is added up described a plurality of bit entropy rate figure, obtains the saliency map of described image or video data.Above-mentioned

Above-mentioned visual attention computing method based on bit entropy rate, in the preferred described bit entropy rate figure obtaining step, described bit entropy rate is used for determining the full connection layout of described each subband feature figure correspondence, each node i is to the average information DER of other nodes _iBe expressed as:

${\mathrm{SER}}_i={\mathrm{\π}}_i\underset{j}{\mathrm{\Σ}}-P_{\mathrm{ij}}\log P_{\mathrm{ij}}$

Wherein, π _iBe the static probability of random walk process, P _IjBe the transition probability of node i to node j, i, j are natural number.

Above-mentioned visual attention computing method based on bit entropy rate in the preferred described saliency map obtaining step, is integrated theory based on feature, and described a plurality of bit entropy rate figure are added up, and obtains the saliency map of described image or video data.

Above-mentioned visual attention computing method based on bit entropy rate, in the preferred described filter step, employing independent component analysis method is respectively gray level image and coloured image is learnt sparse code book.

Above-mentioned visual attention computing method based on bit entropy rate, in the preferred described filter step, adopt described sparse code book basis function that each frame of video is carried out filtering, obtain a plurality of subband feature figure of each frame, j subband characteristic pattern of t frame, upgrade according to following formula:

${f^{\′}}_j(x,y,t)=|f_j(x,y,t)-\underset{\mathrm{\τ}=1}{\overset{k}{\mathrm{\Σ}}}\exp (-\frac{\mathrm{\τ}}{\mathrm{\σ}})f_j(x,y,t-\mathrm{\τ})|$

Wherein, f _j(x, y t) are j subband characteristic pattern of the t frame before upgrading, f ' _j(x, y are j subband characteristic patterns of the t frame after upgrading t), and σ is the characteristic decay rate, and σ is set to 1.5, utilize t frame k frame before to come the t frame is upgraded.

On the other hand, the present invention also provides a kind of vision attention computing system based on bit entropy rate, comprising: filtration module, full connection layout are set up module, bit entropy rate figure acquisition module and saliency map acquisition module.Wherein, filtration module is used to adopt sparse code book basis function that image or video data are carried out filtering, obtains a plurality of subband feature figure, and wherein said sparse code book basis function is based on the sparse code book of study and obtains; Full connection layout is set up each subband feature figure that module is used to be respectively described a plurality of subband feature figure and is set up corresponding full connection layout; Bit entropy rate figure acquisition module is used for adopting on each described full connection layout the method for random walk to carry out the transmission of information, in the random walk process, carries out significance tolerance according to bit entropy rate, and then obtains the bit entropy rate figure of each described subband feature figure correspondence; The saliency map acquisition module is used for described a plurality of bit entropy rate figure are added up, and obtains the saliency map of described image or video data.

Above-mentioned vision attention computing system based on bit entropy rate, in the preferred described bit entropy rate figure acquisition module, described bit entropy rate is used for determining the full connection layout of described each subband feature figure correspondence, each node i is to the average information of other nodes, described bit entropy rate SER _iBe expressed as:

$\mathrm{SE}{R}_i={\mathrm{\π}}_i\underset{j}{\mathrm{\Σ}}-P_{\mathrm{ij}}\log P_{\mathrm{ij}}$

Wherein, π _iBe the static probability of random walk process, P _IjBe the transition probability of node i to node j.

Above-mentioned vision attention computing system based on bit entropy rate in the preferred described saliency map acquisition module, is integrated theory based on feature, and described a plurality of bit entropy rate figure are added up, and obtains the saliency map of described image or video data.

Above-mentioned vision attention computing system based on bit entropy rate, in the preferred described filtration module, employing independent component analysis method is respectively gray level image and coloured image is learnt sparse code book.

Above-mentioned vision attention computing system based on bit entropy rate in the preferred described filtration module, adopts described sparse code book basis function that video data is carried out filtering, obtain a plurality of subband feature figure after, j subband characteristic pattern of t frame, upgrade according to following formula:

${f^{\′}}_j(x,y,t)=|f_j(x,y,t)-\underset{\mathrm{\τ}=1}{\overset{k}{\mathrm{\Σ}}}\exp (-\frac{\mathrm{\τ}}{\mathrm{\σ}})f_j(x,y,t-\mathrm{\τ})|$

Wherein, f _j(x, y t) are j subband characteristic pattern of the t frame before upgrading, f ' _j(x, y are j subband characteristic patterns of the t frame after upgrading t), and σ is the characteristic decay rate, and σ is set to 1.5, utilize t frame k frame before to come the t frame is upgraded.

In prior art, the present invention has following several advantage:

The first, no matter passed through a large amount of experiment showed, is in analysis of image significance or the analysis of video significance, all more accurate than existing additive method based on the analysis result that the present invention obtains.

The second, the computing method of Ti Chuing have the foundation of physiology, psychology aspect to support: the sparse coding characteristic of simple cell instructs this model to adopt sparse code book to generate the subband feature figure of image in the initial stage visual cortex, the regeneration that exists in the visual cortex neural network is local to be connected and the horizontal connected mode of long scope instructs this model to adopt a kind of full graph model that is connected as basic expression on each subband feature figure, and neuronic behavior is driven by total cynapse input of peripheral nerve unit instructs us to propose bit entropy rate.

Three, simultaneously can be from this model of model around two kinds of type of drive-information maximization of vision attention and the center-explain.

Description of drawings

Fig. 1 is the flow chart of steps that the present invention is based on the visual attention computing method embodiment of bit entropy rate;

Fig. 2 is the processing procedure synoptic diagram that the present invention is based on the visual attention computing method of bit entropy rate;

Fig. 3 is the sparse code book synoptic diagram about gray level image that extracts among the present invention;

Fig. 4 A predicting the outcome that be the visual attention computing method that the present invention is based on bit entropy rate to the vision attention point of 3 width of cloth coloured images;

Fig. 4 B predicting the outcome that be the visual attention computing method that the present invention is based on bit entropy rate to the vision attention point of other 3 width of cloth coloured images;

Fig. 5 is the structured flowchart that the present invention is based on the vision attention computing system embodiment of bit entropy rate.

Embodiment

For above-mentioned purpose of the present invention, feature and advantage can be become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments.

With reference to Fig. 1, Fig. 1 is the flow chart of steps that the present invention is based on the visual attention computing method embodiment of bit entropy rate, comprises the steps:

Filter step 110 is learnt sparse code book, obtains sparse code book basis function; Adopt described sparse code book basis function that image or video data are carried out filtering, obtain a plurality of subband feature figure.

Full connection layout establishment step 120 is respectively described each subband feature figure and sets up corresponding full connection layout.

Bit entropy rate figure obtaining step 130 adopts the method for random walk to carry out the transmission of information on described each full connection layout, in the random walk process, makes to carry out significance tolerance according to bit entropy rate, and then obtains the bit entropy rate figure of each described subband feature figure correspondence.

Saliency map obtaining step 140 is added up described a plurality of bit entropy rate figure, obtains the saliency map of described image or video data.

With reference to Fig. 2, Fig. 2 is the processing procedure synoptic diagram that the present invention is based on the visual attention computing method of bit entropy rate.Among Fig. 2, a is the image or the video data of input.At first the sparse code book basis function b that obtains with study carries out filtering to input picture or video data a, the corresponding sparse code book of each the subband feature figure c that obtains; For having set up a full connection layout d, each subband feature figure c represents that this full connection layout d represents to describe the relation of pixel far away in the image then; For the transmission of signal (information) between each neuron in the simulative neural network, this method has adopted the method for random walk on the full connection layout d of each correspondence, entropy rate (Entropy Rate) has been described the average information of random walk process, this total average information is distributed to each figure node to get on, a new significance tolerance-bit entropy rate has been proposed, it has also described the average information of each node to every other node simultaneously, just can obtain a bit entropy rate figure e corresponding to each subband feature figure like this; At last, all bit entropy rate figure are added up just obtained saliency map f.

Below in conjunction with Fig. 3 and Fig. 4 above-mentioned each step is elaborated.

In the filter step 110, after the sparse code book study, obtain sparse code book basis function, adopt this sparse code book basis function that target image or video are carried out filtering, below sparse code book is described.

A large amount of evidences show to have only a spot of early vision neuron to be activated when stimulation appears at the cell receptive field, in order to simulate this specific character, the sparse coding theory is suggested the immanent structure of the expression natural image of making a return journey.Sparse code book basis function is V _k, k is the index to basis function position, direction and yardstick, then image I can be expressed as

I＝∑ _kα _kB _k

α _kBe the coefficient of basis function, this method adopts α _kAs the early vision feature, it can be by the filter function G corresponding with basis function _kCalculate:

α _k＝∑ _x，yG _k(x，y)I(x，y)

Present embodiment adopts independent component analysis method (Independent ComponentAnalysis, ICA) be respectively gray level image and coloured image and learnt the sparse code book of a cover, the ICA here is a kind of method that mixed signal is resolved into subsignal separate on the statistical significance.As shown in Figure 3, the sparse code book synoptic diagram for extracting among the present invention about gray level image.

Full connection layout establishment step 120 is set up corresponding full connection layout for being respectively described each subband feature figure, specifies as follows:

Corresponding to each subband feature figure F _kSet up a full connection layout G _K={ V _k, E _k, V wherein _k={ v _K1..., v _KnBe node corresponding to the image pixel place, v _Ki=(x _i, y _i, f _k(x _i, y _i)) position and two attributes of characteristic response, E arranged _k={ e _Kij, i, j=1 ... n} is the weighting limit between the node, wherein e _Kij=(i, j, w _Kij).Weight w _KijComprise feature difference degree Φ _KijWith space length d _IjThese two

w _kij＝Φ _kij*d _ij

Φ wherein _KijAnd d _IjCan be expressed as

Φ _kij＝exp{|f _k(x _i，y _i)-f _k(x _j，y _j)|/M _k}

$d_{\mathrm{ij}}=\exp \{-\mathrm{\λ}\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2}/D\}$

M in this model _kBe that maximum characteristic response among each subband feature figure is poor, D is the maximum dimension (being the maximal value in width and the height) of image, and λ is used for regulating this importance of two, is traditionally arranged to be 5.

Bit entropy rate figure obtaining step 130 illustrates the bit entropy rate (SiteEntropy Rate) that is used to measure significance.

For the transmission of signal between the neuron in the simulative neural network, on the pairing full connection layout of each subband feature figure, implemented a random walk process (Random Walk), transition probability from node i to node j in this random walk process be (below in order to describe the index k that simply dispenses subband feature figure, as w _KijBe w _Ij)

$P_{\mathrm{ij}}=\frac{w_{\mathrm{ij}}}{{\mathrm{\Σ}}_j{w}_{\mathrm{ij}}}$

The entropy rate of random walk is the total average information that is used for describing this stochastic process, it has also described total quantity of information that all figure nodes (neuron) transmit simultaneously, the entropy rate is decomposed each node to get on, this model just proposed a new notion-bit entropy rate (Site Entropy Rate, SER)-go to describe the average information of each node to other all nodes

${\mathrm{SER}}_i={\mathrm{\π}}_i\underset{j}{\mathrm{\Σ}}-P_{\mathrm{ij}}\log P_{\mathrm{ij}}$

π wherein _iIt is the static probability of random walk process.Obtain maximized principle according to human visual system's information, this model proposes to describe with SER the visual saliency of each node.

At last, integrate theoretical (Feature-IntegrationTheory based on the feature that Treisman proposes, be that vision system at first can extract initial stage visual signature formation characteristic pattern, characteristic pattern is integrated into the focus that a remarkable figure goes to instruct people then), significantly figure obtains (recovering subband feature index of the picture k here again) by the pairing SER figure of each characteristic pattern phase Calais:

$S_i=\underset{k}{\mathrm{\Σ}}{\mathrm{SER}}_{\mathrm{ki}}$

S wherein _iIt is the remarkable value of node i.

With reference to Fig. 4 A and Fig. 4 B, Fig. 4 A is the visual attention computing method that the present invention is based on bit entropy rate to the predicting the outcome of the vision attention point of 3 width of cloth coloured images, Fig. 4 B predicting the outcome to the vision attention point of other 3 width of cloth coloured images that be the visual attention computing method that the present invention is based on bit entropy rate.In Fig. 4 A and Fig. 4 B, one of the leftmost side is classified the original image of input as, middle one classifies the saliency map that adopts method provided by the present invention to obtain as, and rightmost one classifies the lime light data (eye movement data) that people are recorded by the eye movement instrument as when the observation original image.As can be seen, saliency map and eye movement data are very close, and it is more accurate that this method that has illustrated that the present invention proposes is predicted the human eye lime light.

Saliency map obtaining step 140 evidence show in Neuscience: in signal transduction process, have only that unpredictable signal just can be delivered to next stage in present stage; Simultaneously electrophysiology evidence show: neural response can along with cellular exposure under same stimulation time span and sharply descend.Based on these facts, this model thinks and if the node place produces the signal at unpredictable signal or node place variation has taken place that this node is exactly relatively more significant.In video is noted analyzing, when calculating t remarkable figure constantly, should balance out t constantly before the influence of frame of video, specifically, the weighting subband feature figure of number frame of need deducting over upgrades the subband feature figure of present frame.f _j(x, y t) are j subband characteristic pattern of t frame, upgrade it with following formula:

${f^{\′}}_j(x,y,t)=|f_j(x,y,t)-\underset{\mathrm{\τ}=1}{\overset{k}{\mathrm{\Σ}}}\exp (-\frac{\mathrm{\τ}}{\mathrm{\σ}})f_j(x,y,t-\mathrm{\τ})|$

Wherein, f ' _j(x, y are j subband characteristic patterns of the t frame after upgrading t), and σ is the characteristic decay rate, and σ is set to 1.5, utilize t frame k frame before to come the t frame is upgraded.The full connection layout of follow-up foundation is noted analyzing consistent with calculating bit entropy rate and last saliency map with image.

With reference to Fig. 5, Fig. 5 is the structured flowchart that the present invention is based on the vision attention computing system embodiment of bit entropy rate, comprises the steps:

Filtration module 52 is used to learn sparse code book, obtains sparse code book basis function; Adopt described sparse code book basis function that image or video data are carried out filtering, obtain a plurality of subband feature figure.

Full connection layout is set up module 54, is used to be respectively described each subband feature figure and sets up corresponding full connection layout.

Bit entropy rate figure acquisition module 56, be used on described each full connection layout, adopting the method for random walk to carry out the transmission of information, in the random walk process, make to carry out significance tolerance according to bit entropy rate, and then obtain the bit entropy rate figure of each described subband feature figure correspondence.

Saliency map acquisition module 58 is used for described a plurality of bit entropy rate figure are added up, and obtains the saliency map of described image or video data.

Wherein, in the bit entropy rate figure acquisition module 54, bit entropy rate is used for determining the full connection layout of described each subband feature figure correspondence, and each node i is to the average information of other nodes, and described bit entropy rate is expressed as:

${\mathrm{SER}}_i={\mathrm{\π}}_i\underset{j}{\mathrm{\Σ}}-P_{\mathrm{ij}}\log P_{\mathrm{ij}}$

π _iBe the static probability of random walk process, P _IjBe the transition probability of node i to node j.

In saliency map acquisition module 58, can integrate theory based on feature, described a plurality of bit entropy rate figure are added up, obtain the saliency map of described image or video data.

In described filtration module 52, employing independent component analysis method is respectively gray level image and coloured image is learnt sparse code book.

In addition, native system has also proposed the update strategy of the subband feature figure of present frame during video is noted analyzing.Specify as follows:

In the filtration module, adopt described sparse code book basis function that video data is carried out filtering, obtain a plurality of subband feature figure after, j subband characteristic pattern of t frame, upgrade according to following formula:

${f^{\′}}_j(x,y,t)=|f_j(x,y,t)-\underset{\mathrm{\τ}=1}{\overset{k}{\mathrm{\Σ}}}\exp (-\frac{\mathrm{\τ}}{\mathrm{\σ}})f_j(x,y,t-\mathrm{\τ})|$

Wherein, f _j(x, y t) are j subband characteristic pattern of the t frame before upgrading, f ' _j(x, y are j subband characteristic patterns of the t frame after upgrading t), and σ is the characteristic decay rate, and σ is set to 1.5, utilize t frame k frame before to come the t frame is upgraded.

Need to prove that the foregoing description is similar based on the vision attention computing system of bit entropy rate with visual attention computing method principle based on bit entropy rate, relevant part can be with reference to the explanation for method embodiment.Do not repeat them here.

More than a kind of visual attention computing method and system based on bit entropy rate provided by the present invention is described in detail, used specific case herein principle of the present invention and embodiment are set forth, the explanation of above embodiment just is used for helping to understand method of the present invention and core concept thereof; Simultaneously, for one of ordinary skill in the art, according to thought of the present invention, the part that all can change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention.

Claims (6)

1. the visual attention computing method based on bit entropy rate is characterized in that, comprises the steps:

Filter step adopts sparse code book basis function that image or video data are carried out filtering, obtains a plurality of subband feature figure, and wherein said sparse code book basis function is based on the sparse code book of study and obtains;

Full connection layout establishment step, each the subband feature figure that is respectively described a plurality of subband feature figure sets up corresponding full connection layout;

Bit entropy rate figure obtaining step adopts the method for random walk to carry out the transmission of information on each described full connection layout, in the random walk process, carries out significance tolerance according to bit entropy rate, and then obtains the bit entropy rate figure of each described subband feature figure correspondence;

The saliency map obtaining step is added up described a plurality of bit entropy rate figure, obtains the saliency map of described image or video data; And

In the described filter step, adopt described sparse code book basis function that each frame of video is carried out filtering, obtain a plurality of subband feature figure of each frame, j subband characteristic pattern of t frame, upgrade according to following formula:

Wherein, f _j(x, y t) are j subband characteristic pattern of the t frame before upgrading, f ' _j(x, y are j subband characteristic patterns of the t frame after upgrading t), and σ is the characteristic decay rate, and σ is set to 1.5, utilize t frame k frame before to come the t frame is upgraded;

In the described full connection layout establishment step, the full connection layout of each subband feature figure correspondence is set up according to following mode:

Set up a full connection layout G corresponding to each subband feature figure Fk _K={ V _k, E _k, V wherein _k={ v _K1..., v _KnBe node corresponding to the image pixel place, v _Ki=(x _i, y _i, f _k(x _i, y _i)) position and two attributes of characteristic response, E arranged _k={ e _Kij, i, j=1 ... n} is the weighting limit between the node, wherein e _Kij=(i, j, w _Kij); Weight w _KijComprise feature difference degree Φ _KijWith space length d _IjThese two,

w _kij＝Φ _kij*d _ij

Φ wherein _KijAnd d _IjBe expressed as

Φ _kij＝exp{|f _k(x _i，y _i)-f _k(x _j，y _j)|/M _k}

M _kBe that maximum characteristic response among each subband feature figure is poor, D is the maximum dimension of image, and λ is used for regulating this importance of two, is set to 5;

In the described bit entropy rate figure obtaining step, described bit entropy rate is used for determining the full connection layout of described each subband feature figure correspondence, and each node i is to the average information SER of other nodes _iBe expressed as:

Wherein, π _iBe the static probability of random walk process, P _IjBe the transition probability that node i arrives node j, i, j are natural number.

2. the visual attention computing method based on bit entropy rate according to claim 1, it is characterized in that, in the described saliency map obtaining step, integrate theoretical based on feature, described a plurality of bit entropy rate figure are added up, obtain the saliency map of described image or video data.

3. the visual attention computing method based on bit entropy rate according to claim 2 is characterized in that, in the described filter step, employing independent component analysis method is respectively gray level image and coloured image is learnt sparse code book.

4. the vision attention computing system based on bit entropy rate is characterized in that, comprising:

Filtration module is used to adopt sparse code book basis function that image or video data are carried out filtering, obtains a plurality of subband feature figure, and wherein said sparse code book basis function is based on the sparse code book of study and obtains;

Full connection layout is set up module, and each the subband feature figure that is used to be respectively described a plurality of subband feature figure sets up corresponding full connection layout;

Bit entropy rate figure acquisition module is used for adopting on each described full connection layout the method for random walk to carry out the transmission of information, in the random walk process, carries out significance tolerance according to bit entropy rate, and then obtains the bit entropy rate figure of each described subband feature figure correspondence;

The saliency map acquisition module is used for described a plurality of bit entropy rate figure are added up, and obtains the saliency map of described image or video data; Wherein

In the described filtration module, adopt described sparse code book basis function that each frame of video is carried out filtering, obtain a plurality of subband feature figure of each frame, j subband characteristic pattern of t frame, upgrade according to following formula:

Wherein, f _j(x, y t) are j subband characteristic pattern of the t frame before upgrading, f ' _j(x, y are j subband characteristic patterns of the t frame after upgrading t), and σ is the characteristic decay rate, and σ is set to 1.5, utilize t frame k frame before to come the t frame is upgraded;

Described full connection layout is set up in the module, and the full connection layout of each subband feature figure correspondence is set up according to following mode:

Corresponding to each subband feature figure F _kSet up a full connection layout G _K={ V _k, E _k, V wherein _k={ v _K1..., v _KnBe node corresponding to the image pixel place, v _Ki=(x _i, y _i, f _k(x _i, y _i)) position and two attributes of characteristic response, E arranged _k={ e _Kij, i, j=1 ... n} is the weighting limit between the node, wherein e _Kij=(i, j, w _Kij); Weight w _KijComprise feature difference degree Φ _KijWith space length d _IjThese two,

w _kij＝Φ _kij*d _ij

Φ wherein _KijAnd d _IjBe expressed as

Φ _kij＝exp{|f _k(x _i，y _i)-f _k(x _j，y _j)|/M _k}

M _kBe that maximum characteristic response among each subband feature figure is poor, D is the maximum dimension of image, and λ is used for regulating this importance of two, is set to 5;

In the described bit entropy rate figure acquisition module, described bit entropy rate is used for determining the full connection layout of described each subband feature figure correspondence, and each node i is to the average information SER of other nodes _iBe expressed as:

Wherein, π _iBe the static probability of random walk process, P _IjBe the transition probability that node i arrives node j, i, j are natural number.

5. the vision attention computing system based on bit entropy rate according to claim 4, it is characterized in that, in the described saliency map acquisition module, integrate theoretical based on feature, described a plurality of bit entropy rate figure are added up, obtain the saliency map of described image or video data.

6. the vision attention computing system based on bit entropy rate according to claim 5 is characterized in that, in the described filtration module, employing independent component analysis method is respectively gray level image and coloured image is learnt sparse code book.

Images