CN103218815B - Utilize the method for natural scene statistical computation image saliency map - Google Patents

Abstract

The invention belongs to the technical field of image saliency map models, in particular to a method for statistically calculating image saliency maps using natural scenes. The invention uses the multiplier random variable in the Gaussian scale mixed statistical distribution of the natural scene to calculate the image saliency map, thereby establishing the image saliency map model. The analysis shows that the saliency map model proposed by the present invention has a high consistency with the visual attention selection mechanism, that is, it can highlight the visual stimuli with high salience while suppressing the recurring stimuli, thereby better describing the image pair Saliency distribution of visual stimuli for human eyes.

Description

A Method for Computing Image Saliency Map Using Natural Scene Statistics

technical field

The invention belongs to the technical field of image saliency map models, and in particular relates to a method for calculating image saliency maps by using multiplier random variables in the Gaussian-scale mixed statistical distribution of natural scenes.

Background technique

Visual Attention (VA) is an important mechanism of the Human Visual System (HVS). Generally speaking, the human eye will receive a large number of visual stimuli from the outside world at one moment, but due to the limited processing resources in the HVS, there will be a competition for processing resources between different visual stimuli. In the end, the most informative stimulus among all visual stimuli will win the competition, while other stimuli will be suppressed ^[1-2] . HVS uses such a VA selection mechanism to make full use of effective resources to process a large number of visual stimuli, thereby reducing the complexity of scene analysis ^[3] .

In the fields of neurology and biology, some experiments use specific equipment to record the gaze points of the observers when they face different image scenes, and other experiments allow the observers to actively mark the areas of interest in different image scenes. The purpose of these experiments is to study the VA mechanism of HVS based on the experimental results. The research results show that: on the one hand, places with large grayscale changes in the image will attract more attention from HVS, including texture information, edge information, and so on. On the other hand, HVS has a certain inhibitory effect on repeated redundant information, and novel information that is not repeated with the surroundings will attract more attention from HVS. Therefore, we say that the less the number of occurrences of visual stimulus structure templates, the greater the significance ^[3] .

With the continuous deepening of the research on VA mechanism, some research results have been applied to the problem of image processing, and brought some enlightening results. However, in the real-time processing of images, it is unrealistic to obtain the distribution of fixation points of human eyes on different images through experiments. It is necessary to establish a computable VA model that can simulate the properties of HVS, so as to realize the prediction of the salience of visual stimuli. Among these models, the most basic is the saliency map model, which uses a saliency map to describe the degree of attention of HVS to different positions in the scene ^[1] .

The most basic conclusion of visual attention is the attention feature integration theory proposed by Treisman & Gelade in 1980 on the basis of experiments ^[5] . In 1989, Wolfe et al. proposed the GuidedSearch model ^[6] , which uses saliency maps to search for objects in the scene. In 1985, Koch & Ullman established the VA model framework based on neurological theory ^[7] . The classic Itti&Koch saliency map method ^[8] uses multi-scale and multi-channel to model the VA mechanism in the Koch&Ullman framework, extracts relevant features, and obtains a saliency map with high consistency with HVS through fusion. The STB (SaliencyToolBox) ^[9] method combines the theory of consistent visual cognition to improve the Itti&Koch method. In addition, literature [16] proposed a saliency map model (PFT-based Saliencymap, PFTSal) based on the Fourier transform phase spectrum of the image, and it has been widely recognized ^[1][4][17] . Recently, literature [14] proposed a cosine transform-based PCT (Pulse Discretecosine Transform) model, which was later called the signature image descriptor by literature [15]. Their experimental results confirmed that the image descriptor established on this descriptor The consistency between the signature-based Saliencymap (signatureSal) and HVS is higher than other existing saliency map models. Other works include the Bruce method ^[10] based on information theory, and the SUN model ^[12] and Surprise model ^[13] based on Bayesian methods.

The present invention proposes a statistical saliency graph model for natural scenes. It utilizes multiplier random variables in the Gaussian scale mixture (GSM, Gaussian scale mixture) statistical distribution of natural scenes to compute image saliency maps. The saliency graph model of the present invention has high consistency with the VA mechanism.

Contents of the invention

The object of the present invention is to provide a method for calculating an image saliency map consistent with the attention visual selection mechanism of the human visual system, so as to establish an image saliency map model.

The present invention uses the multiplier random variable in the Gaussian scale mixture (GSM) statistical distribution of natural scenes to calculate the image saliency map, and the specific steps are as follows:

(1) Set the image as a grayscale image , , are the number of rows and columns of the image respectively, and perform wavelet coefficient transformation on it to obtain multiple wavelet coefficient subbands.

(2) In each wavelet coefficient subband, for each coefficient Select a suitable wavelet coefficient neighborhood, and pull the wavelet coefficient neighborhood into a wavelet coefficient neighborhood vector ,in is the neighborhood size;

According to the statistical properties of the Natural Scene Statistics (NSS) model, the wavelet coefficient neighborhood vector of the natural image It can be described by Gaussian scale mixture (GSM) distribution, namely ,in is a random multiplier that characterizes the variation of neighborhood vector covariance, is zero mean, and the covariance matrix is Gaussian random variables of , therefore, the neighborhood vector The probability density function of Expressed as:

(1)

in, is a random multiplier The probability density function distribution of ;

Therefore, the coefficient neighborhood vector About Random Multiplier Obey the zero mean, the covariance matrix is The Gaussian distribution of , the conditional probability distribution function is expressed as:

(6).

(3) Calculate the estimated value of the Gaussian scale mixture distribution multiplier variable

When the selected neighborhood of wavelet coefficients is small enough, that is enough hours, assuming the multiplier remains constant within this neighborhood, so it is possible to temporarily As a definite quantity or constant, at this time, the neighborhood corresponding multiplier The conditional probability distribution function can be The maximum likelihood estimation of is obtained, namely ^[18] :

(7)

remember The eigenvalues of are decomposed into ,in for The eigenvector of The matrix formed, for The eigenvalues of constituted matrix.

therefore, The maximum likelihood estimation result of is:

(8)

without loss of generality ,Have The covariance matrix of . Therefore, you can use The covariance matrix of replace ,get

(9)

superscript of express The transpose of , assuming the wavelet vector is a feature sample, then The set of σ constitutes the feature space of the image. Since the mean value of wavelet coefficients is zero, therefore, for to the center of this feature space Mahalanobis distance ^[22] . According to the nature of the Mahalanobis distance: the Mahalanobis distance comprehensively considers the relationship between the various dimensions of the feature vector, the sample Mahalanobis distance to center of feature space bigger, The lower the probability of belonging to the feature space, that is to say, the sample The higher the "significance" of , and vice versa.

According to the description of formula (9), we can know and There is a proportional relationship, that is

(10 )

therefore, It is an effective description of the significance of the sample in the feature space. The higher the significance of feature samples, the corresponding The larger the value; on the contrary, the lower the significance of the feature sample, the corresponding The smaller the value.

According to the nature of image wavelet decomposition, the wavelet coefficient can extract the discontinuous information of the image, and at the same time, compared with the frequency domain, it can describe the distribution of HVS attention intensity to the scene. If we select all the adjacent coefficients of each wavelet coefficient in its space, scale and direction to form a neighborhood, and express all the coefficients in the neighborhood as a neighborhood vector . So It describes the feature vector of visual stimuli in this neighborhood. Therefore, by The feature space composed of all sets of the visual stimuli not only describes the spatial distribution of visual stimuli, but also describes the correlation information between adjacent visual stimuli in space, scale and direction, such as the adjacent coefficient in the same subband describes the spatial correlation. The correlation of adjacent visual features, the adjacent coefficients of the same scale and different directions describe the correlation of visual features of adjacent directions, and the adjacent coefficients of different scales and the same direction describe the correlation of visual features of adjacent scales. Then from formula (6) and the above analysis, it can be shown that: constituted The matrix can comprehensively describe the salience distribution of visual scenes under the premise of comprehensively considering the spatial distribution of visual stimuli and the correlation between adjacent visual stimuli. Such a comprehensive description is comparable to the VA mechanism described in the fields of neurology and biology. consistent.

(4) By fusing the saliency corresponding to all wavelet coefficient subbands, a complete natural scene statistical saliency map NSSSal (Natural Scene Statistical Saliencymap) model can be obtained:

( 2)

In the formula, Indicates the scale Superposition of saliency descriptions in different directions on ; Represents the addition of different scales (across-scaleaddition) ^[9] , interpolating the saliency descriptions on all scales to post-addition, where It is a Gaussian blur kernel function, which is used to smooth the saliency map ^[15] .

(5) Adjust the dynamic range of the grayscale in formula (11) to , the closer the value is to 1, the more significant the corresponding area in the image is, and it is easier to attract the attention of the human eye than other places with smaller values.

(6) If the image is a RGB modulated color image , with a red channel , green channel with blue channel . Calculate the corresponding grayscale channel , red-green antagonistic pair Antagonized with yellow and blue .

grayscale channel for:

(12)

According to the human eye's processing mechanism for color information, the four generalized modulation red (R), green (G), blue (B) and yellow (Y) channels are:

(13)

(14)

(15)

(16)

red-green antagonistic pair Antagonized with yellow and blue for:

(17)

(18)

(7) For the grayscale channel , red-green antagonistic channel Antagonized with yellow and blue According to the calculations in steps (1)-(5), the obtained values are denoted as , and the saliency map of the color image, the weighted average of these three is taken as the saliency map of the color image ,which is:

(19)

in, , and are the weights of the three channels respectively, and have .

According to the present invention, in the saliency map calculated for the image, positions with higher pixel values correspond to higher saliency of the image; positions with lower pixel values correspond to lower saliency of the image.

The saliency map model proposed by the present invention has a high consistency with the visual attention selection mechanism, that is, it can highlight the visual stimuli with high salience while suppressing the recurring stimuli, thereby better describing the impact of images on human vision. The distribution of salience for stimuli.

Description of drawings

Figure 1: Steerable pyramid decomposition and neighborhood selection of wavelet coefficients.

detailed description

In the following, the present invention compares the performance of the saliency map model NSSSal/cNSSSal of the present invention and other saliency map models for extracting saliency maps from natural images through examples. At the same time, the public Bruce database ^[10] and ImgSal database ^[11] will also be used to quantitatively evaluate their AUC (Area Under the Curve, the area under the ROC curve).

In the experiment, a complete steerable pyramid is used to decompose the image by wavelet to preserve the directionality of the image. Decomposition scale number , the number of directions , respectively , , and ,As shown in Figure 1. At the same time, each wavelet coefficient is selected (marked by small black squares in the figure) and its adjacent coefficients in the same subband, coefficients at the same position in subbands of the same scale and different directions, and coefficients at the same position on the parent scale of the same direction (marked by small gray squares in the figure) together form its neighborhood vector , the neighborhood size . The Gaussian kernel variance is selected to be 0.045 times the width of the saliency map.

For any natural image, the saliency map obtained by a certain saliency map model is . Select threshold , according to the threshold right Perform binarization, and record the binarized for . According to the subjective attention distribution map provided in the database (FixationDensityMap), its positive class rate (TPR, TruePositiveRate) is:

(20)

in, The symbol represents multiplication between pixels, Represents the 1 norm value, that is, the matrix In is the number of 1 elements.

Similarly, its false alarm rate (FPR, FalsePositiveRate) is:

(twenty one)

For a given threshold , will be obtained for all images in the database The mean value of the saliency map model as the threshold is Positive class rate when , similarly, will The average value of as the threshold is false alarm rate . By choosing different thresholds ,by as the vertical axis, with On the horizontal axis, draw the ROC curve. Area under the ROC curve Provides a measure of agreement between saliency map models and the subjective attention distribution of the human eye, The closer the value is to 1, the higher the consistency between the saliency graph model and the VA mechanism.

(1) Performance evaluation of grayscale image saliency map model

Table 1 is calculated by NSSSal, Itti&Koch, PFTSal, and signatureSal models after grayscale images in the Bruce and ImgSal databases value, The closer the value is to 1, the higher the consistency between the saliency map model and the VA mechanism. It can be seen that on these two databases, Itti&Koch's The result is the lowest, and the PFTSal and signatureSal models are statistically similar, while the NSSSal model has a value closer to 1 than the two. value, the highest consistency with the VA mechanism.

Table 1: AUC value comparison of grayscale image saliency map

.

(2) Performance evaluation of color image saliency map model

Further, for color natural images, we compared the ROC (Receiver Operating Characteristic) curves of cNSSSal and Itti&Koch ^[8] , PQFTSal ^[16] , signatureSal ^[15] , and the area under the ROC curve AUC (AreaUndertheCurve). Taken in the experiment , that is, take the same weight for the three channels.

Table 2 is calculated by the cNSSSal, Itti&Koch, PQFTSal, signatureSal models on the Bruce and ImgSal databases value. After considering the color information of different saliency map models, The values will increase accordingly, while the value of cNSSSal is still the highest compared with other models.

Table 2: AUC value comparison of grayscale image saliency map

.

Practical application effect of the present invention

Saliency graph models are widely used in image processing problems such as adaptive image compression, video summarization, coding and image progressive transmission, image segmentation, image and video quality assessment, object recognition, and content-aware image scaling. The following takes the application of the saliency graph model of the present invention in image quality assessment as an example to illustrate the effectiveness and superiority of the saliency graph model of the present invention.

Image quality assessment measures simulate the overall mechanism of the human visual system, hoping to obtain results that are consistent with the human eye's assessment of image quality. However, the important attentional visual selection mechanism in HVS is often neglected in many existing image quality assessment methods. When ignoring the attentional visual selection mechanism, image quality assessment measures assume that the human eye attaches equal attention to all objects including natural scenes and image distortions. However, according to the attentional visual selection mechanism, the human visual model does not view images in the form of simple high-dimensional spatial signals, but has different sensitivities to different attributes of images, such as brightness, contrast, shape and texture of objects, and orientation , smoothness, etc. Since the human visual system has different sensitivities to different components of an image, such different sensitivities need to be taken into account when evaluating the quality of distorted images. Therefore, combining the attention visual selection mechanism can effectively improve the performance of image quality assessment algorithms.

The image saliency map that will be obtained in the experiment As a weight, to readjust the weight ratio of each part of the quality assessment results in the final result in the original image quality assessment method, the purpose is to make the assessment results more consistent with the subjective assessment results of the human eye. The experiments weight three objective image quality assessment measures (PSNR, MSSIM ^[25] , VIF ^[26] ) with saliency maps.

(1) PSNR based on saliency map

Since PSNR is based on pixels, the saliency map can be directly used as a weighting matrix. suppose is a saliency map on the first gray value of a pixel. The PSNR (Saliency-basedPSNR, SPSNR) based on the saliency map is defined as:

(twenty two)

in, is the total number of pixels in the image, , image respectively , on the first The gray value of each pixel, is the grayscale dynamic range of the image. Since the gray levels of the saliency maps obtained from different natural images are different, the Normalize the weights.

(2) MSSIM based on saliency map

Since MSSIM is based on image sub-blocks, firstly the obtained saliency map It is divided into sub-blocks with the same number and size as the reference image. suppose is divided into sub-blocks, and with sub-blocks , The corresponding sub-block is , the subblock The normalized gray value of As the weight of the corresponding sub-block, then there is

(twenty three)

in, express the number of pixels in the express in the first gray value of a pixel. Therefore, the saliency-based MSSIM (Saliency-basedMSSIM, SMSSIM) can be defined as

(twenty three)

in, , are the reference image and the distorted image respectively block, is the total number of image subblocks.

(3) VIF based on saliency map

Since the spatial domain VIF is multi-scale, the reference image is adjusted to different scales, and the saliency map at different scales is calculated ,in is the scale index. Then according to the nature of the VIF based on the sub-block, in the scale down, put divided into block sub-block is . Subblock The normalized gray value of As the weight of the corresponding sub-block, then there is

(25)

in, express the number of pixels in the express in the first gray value of a pixel. Therefore, the VIF (Saliency-basedVIF, SVIF) based on the saliency map can be defined as

(26)

in, represents the number of scales, is the scale index, Indicates the first The number of image sub-blocks under the scale, index for subblocks. for the reference image the first scale block, , are the sub-blocks of the reference image and the distorted image received by the human eye, are the corresponding model parameters. for and about mutual information.

In the experiment, four different saliency graph models (Itti&Koch ^[8] , PQFTSal ^[16] , signatureSal ^[15] , cNSSSal) were used to weight PSNR, MSSIM, and VIF, and each measure derived five different implementation methods (IQA, Itti_IQA, PQFT_IQA, signature_IQA, and cNSS_IQA).

The matlab implementation programs of MSSIM and spatial domain VIF are all from the implementation version of the original author on the Internet. In the experiment, when calculating MSSIM and SMSSIM, the sub-block size is 8×8, and the distance between adjacent sub-blocks is 1 pixel; while calculating the spatial domain VIF, 4 scales are used, the sub-block size is 3×3, and the sub-blocks are different from each other. overlapping. At the same time, in order to ensure the fairness of the comparison, the parameter settings when using the three models to extract saliency maps and combine them with image quality assessment are consistent. The default fuzzy parameters in the version implemented by the author of signatureSal (see [15] for details) are used to blur the three saliency maps.

The performance of the present invention is verified on the LIVE image quality assessment database ^[28] . The LIVE database contains 982 images, 779 of which are distorted images. These images are obtained from 29 reference images at different distortion levels by JPEG, JPEG2000, white noise, Gaussian blur and channel fast fading. The database also gives the subjective evaluation score (DMOS) corresponding to each image. The range of DMOS is [0,100]. DMOS=0 means that the image is undistorted. As the degree of image distortion increases, the value of DMOS will increase accordingly. By comparing the evaluation results obtained by the DMOS with the image quality evaluation algorithm, the performance of the image quality evaluation algorithm can be evaluated.

After nonlinear regression fitting of image quality estimation results and DMOS, five objective evaluation indicators can be used to quantitatively evaluate the consistency of objective image quality evaluation measures and subjective quality evaluation results: 1) Linear Correlation Coefficient (LCC) , which describes the prediction accuracy, the closer its value is to 1, the higher the prediction accuracy; 2) Mean Absolute Error (MeanAbsoluteError, MAE), the smaller the value, the smaller the prediction absolute error; 3) Root mean square error (RootMeanSquaredError, RMSE), the smaller the value, the smaller the prediction root mean square error; 4) Outlier ratio (OutlierRatio, OR), which describes the prediction consistency, the smaller the value, the higher the prediction consistency; 5) Spearman's Rank Ordered Correlation Coefficient (SROCC), which describes the monotonicity of prediction, the closer its value is to 1, the better the monotonicity of prediction. The quantitative evaluation results are as follows:

Table 3: Comparison of performance improvement of color image saliency map model on image quality assessment measurement

.

The values in bold in Table 3 are the optimal results obtained by several different implementation methods in each evaluation index. It can be seen that three different saliency graph models (Itti&Koch, PQFTSal, signatureSal and cNSSSal) have all improved the performance of image quality assessment measures. At the same time, the cNSSSal saliency map improves the measurement performance of image quality assessment far better than other saliency map models, which proves the superiority of the saliency map model of the present invention in image quality assessment problems.

references

[1] UEngelke, HKaprykowsk, H-JZepernick, and PNdjiki-Nya.Visual attentioninqualityassessment[J].IEEESignalProcessingMagazine,2011,28(6):50-59.

[2]EKowler.Eyemovements:Thepast25years[J].VisionRes.,2011,51(13):1457-1583.

[3]MCarrasco.Visual Attention:Thepast25years[J].VisionRes.,2011,51(13):1484-1525.

[4] AToet.Computationalversuspsychophysicalbottom-upimagesaliency:acomparativeevaluationsstudy[J].IEEETransactions,2011,PAMI-33(11):2131-2146.

[5] AM Treisman and G Gelade. Afeature-integration theory of attention [J]. Cogn. Psychol., 1980, 12(1): 97-136.

[6] JMWolfe, KRCave, and SLFranzel. Guided search: An Alternative to the feature integration model for visual search [J]. J. Exp. Psychol. Hum. Percept. Perform., 1989, 15(3): 419-433.

[7] CKoch and SUllman. Shifts in selective visual attention: towards the underlying neural circuit [J]. Human Neurobiol, 1985, 219-227.

[8] LItti, C Kochand ENiebur. A model of saliency-based visual attention for rapids scene analysis [J]. IEEE Transactions, 1998, PAMI-20 (11): 1254-1259.

[9] DWaltherandCKoch.Modelingattentiontosalientproto-objects[J].NeuralNetworks,2006,19(9):1395-1407.

[10] NDBBruceandJKTsotsos.Saliencybasedoninformationmaximization[A].InProc.AdvancesinNeuralInformationProcessingSystem[C].2005.155-162.

[11] JLi, MD Levine, XAn, XXu, and HHe. Visual saliency based on scale-space analysis in the frequency domain [J]. IEEE Transactions, 2007, PAMI-PP(99): 1,0.

[12] CKanan, MHTong, LZhang, and GW Cottrel. SUN: Top-downsaliency using natural statistics [J]. Visual Cognition, 2009, 17(6/7): 979-1003.

[13] LIttian and P Baldi. Bayesian surprise attract human attention [A]. InProc. Advances in Neural Information Processing System [C]. 2009.547-554.

[14] YYing, BWang, LM Zhang. Pulsed discretecosine transform for saliency-based attention [A]. InProc. IEEEICDL'09 [C]. 2009.1-6, 5-7.

[15] XHou, JHarel, and CKoch. Image signature: highlighting sparse alient regions [J]. IEEE Transactions, 2012, PAMI-34(1): 194-201.

[16]MQiandLMZhang.Saliency-basedimagequalityassessmentcriterion[A].InProc.AdavancedIntelligentComputingTheoriesandApplications[C].2008.1124-1133.

[17] CGuoandLMZhang.Spatio-temporalsaliencydetectionusingphasespectrumofquaternionFouriertransform[A].InProc.IEEECVPR’08[C].2008.1-8,23-28.

[18] MJ Wainwright and EPSimoncelli. Scale mixtures of Gaussians and the statistics of natural images [J]. Adv. Neural Inf. Process. Syst., 2000, 12:855-861.

[19] J Portilla, V Strela, MJ Wainwright, and EPS Simoncelli, Image denoising using scale mixture of Gaussians in the wavelet domain [J]. IEEE Transactions, 2003, Image Processing-12(11): 1338-1351.

[20] ZWang and AC Bovik. Reduced-andno-reference image quality assessment [J]. IEEE, 2011, Signal Processing Magazine-28(6): 29-40.

[21] AK Moorthy and AC Bovik. Statistics of natural image distortions [A]. InProc. IEEE ICASSP'10 [C]. 2010.962-965, 14-19.

[22] PC Mahalanobis. On the generalized distance statistics [J]. Proceeding of the National Institute of Sciences of India, 1936, 2(1): 49-55.

[23] EPSimoncelli, WT Freeman, EHAdelson, and DJ Heeger. Shiftable multiscale transforms [J]. IEEE Transactions, 1992, Informaiton Theory-38(2):587–607.

[24]ZWangandACBovik.Meansquarederror:loveitorleaveit?Anewlookatsignalfidelitymeasures[J].IEEESignalProcessingMagazine,2009,26(1):98-117.

[25] ZWang, AC Bovik, HR Sheikh, and EPSimoncelli. Image quality assessment: From error measurement to structural similarity [J]. IEEE Transactions, 2004, Image Processing-13(4): 600-612.

[26] HR Sheikhand AC Bovik. Image information and visual quality [J]. IEEE Transactions, 2006, Image Processing-15(2): 430-444.

[27] HR. Sheikh, MFSabir, and AC Bovik. Astatistical evaluation of recent full reference image quality assessment algorithms [J]. IEEE Transactions, 2006, Image Processing-15 (11): 3440-2451.

[28] HR Sheikh, Z Wang, AC Bovik, and LK Cormack. Image and Video Quality Assessment Research at LIVE [Online]. Available: http://live.ece.utexas.edu/research/quality.

Claims (2)

1. A method for calculating an image saliency map, characterized in that the multiplier random variable in the natural scene Gaussian scale mixed statistical distribution is used to calculate the image saliency map, and the specific steps are as follows: (1) Set the image as a grayscale image R and C are the number of rows and columns of the image respectively, and the wavelet coefficient transformation is performed on it to obtain a plurality of wavelet coefficient subbands; (2) In each wavelet coefficient subband, select a suitable wavelet coefficient neighborhood for each coefficient c, and pull the wavelet coefficient neighborhood into a wavelet coefficient neighborhood vector Among them, M is the size of the neighborhood; according to the statistical properties of the statistical model of natural scenes, the wavelet coefficient neighborhood vector of the natural image is described by Gaussian scale mixture distribution That is, c'=su', where s is a random multiplier representing the change of the neighborhood vector covariance, _u ' is a Gaussian random variable with zero mean and covariance matrix Cu; (3) Calculate the maximum likelihood estimation result of Gaussian scale mixed distribution multiplier variable as follows: $\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; =</span>\sqrt{{\mathrm{<span\; class="notranslate">\; c</span>}}^{<span\; class="notranslate">\; \&prime;</span>\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; c</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$ Assuming E{s ² }=1, replacing C _u with the covariance matrix C _c of c′, we get: $\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; =</span>\sqrt{{\mathrm{<span\; class="notranslate">\; c</span>}}^{<span\; class="notranslate">\; \&prime;</span>\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; c</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$ Assuming that the wavelet coefficient neighborhood vector c' is a feature sample, then the set of c' constitutes the feature space of the image; Mahalanobis distance from c′ There is a proportional relationship, namely: is an effective description of the significance of the sample in the feature space, the higher the significance of the feature sample, the corresponding The larger the value; on the contrary, the lower the significance of the feature sample, the corresponding The smaller the value; (4) By fusing the saliency corresponding to all wavelet coefficient subbands, a complete natural scene statistical saliency map NSSSal model can be obtained: In the formula, Indicates the scale Superposition of saliency descriptions in different directions on ; Indicates the addition of different scales, interpolating the saliency descriptions on all scales to R×C and adding them, where It is a Gaussian blur kernel function, which is used to smooth the saliency map; (5) Adjust the grayscale dynamic range of formula (4) to [0, 1]. The closer the value is to 1, the more significant the corresponding area in the image is, and it is easier to attract people than places with smaller values. eye attention (6) If the image is a RGB modulated color image The gray channel I of the image, the red-green antagonistic pair channel RG and the yellow-blue antagonistic pair BY are calculated according to steps (1)-(5) to obtain the saliency maps respectively recorded as NSSSal_I, NSSSal_RG and NSSSal_BY, and the weighting of these three averaged as a saliency map for this color image, namely: cNSSSal = ω _I NSSSal_I + ω _RG NSSSal_RG + ω _BY NSSSal_BY (5) Wherein, ω _I , ω _RG and ω _BY are the weights of the three channels respectively, and ω _I +ω _RG +ω _BY =1. 2. The method according to claim 1, in the saliency map calculated for each image, a position with a higher pixel value corresponds to a higher saliency of the image; a position with a lower pixel value corresponds to a lower saliency of the image.