CN107992874B - Method and system for image salient target region extraction based on iterative sparse representation - Google Patents

Abstract

The invention provides an image salient target area extraction method and system based on iterative sparse representation. First, the original image is divided into superpixels by using multiple sets of SLIC segmentation methods with different pixel number parameters, and a group of segmentations with different sizes of superpixel areas are generated. Then, for the segmentation results of each scale, the classical visual attention detection results are used as the initial saliency map to constrain the selection of foreground and background sample regions, and then the reconstruction residuals of each superpixel region are calculated through the sparse representation process as the saliency map. The saliency factor is combined with recursive iterative operation to optimize the saliency detection result map under a single scale, and finally the final salient target and detection result are obtained through multi-scale saliency map fusion. The invention can effectively improve the problems existing in the traditional method, such as inconsistent saliency evaluation of a single target, difficulty in detecting salient targets at the edge of an image, and incomplete extraction of multiple salient targets.

Description

Method and system for image salient target region extraction based on iterative sparse representation

technical field

The invention belongs to the field of computer vision and image processing, and relates to an image salient target area extraction technology based on iterative sparse representation.

Background technique

Image visual saliency analysis is a basic research project that attaches great importance to computer vision, psychology, neuroscience and other fields. reflect. Through image saliency analysis, the target area that people are interested in can be effectively extracted, data compression can be successfully achieved, and the efficient management and utilization of data can be completed, which is also the basic link of many image processing problems.

Since 1998 when people first realized the automatic saliency analysis of images by computer, with the continuous exploration of its application prospects, novel automatic salient target detection algorithms have emerged one after another. From the perspective of solutions, the existing salient target extraction algorithms can be roughly divided into two categories: data-driven bottom-up detection methods and task-driven top-down detection methods. The former automatically processes and recognizes the input image based on empirical knowledge, and realizes the saliency analysis in traditional cognition, which is usually an automatic extraction algorithm in unsupervised mode; The target object that meets the specific application requirements is usually identified as a recognition algorithm under supervised learning. On the other hand, from the perspective of extracting the result state, traditional methods can be divided into saliency analysis algorithms based on visual attention and salient target extraction methods. The former generates pixel-level saliency prediction maps, while the latter extracts complete The salient target area is the final target.

In bottom-up unsupervised methods, due to the lack of high-level biocognitive information, it is usually necessary to introduce certain hypothetical constraints to complete the detection task. The empirical analysis results show that the target distributed near the middle of the image can attract more visual attention, whereas the area close to the edge of the image is usually less saliency. , saliency detection methods combined with image center/boundary constraints and contrast analysis developed rapidly, and also showed very outstanding detection performance. With the deepening and expansion of research and applications, the above-mentioned method's dependence on hypothetical conditions has become increasingly prominent. The specific manifestations are as follows: 1) When the salient target is close to the edge of the image, it is usually impossible to achieve correct detection; 2) Based on local With the contrast analysis method, the extracted salient target area is incomplete, and the saliency evaluation within the target is not uniform; 3) The method based on the global contrast analysis, when dealing with the problem of multiple salient targets at the same time, there are often missed detections. . Therefore, how to overcome the shortcomings of traditional methods, weaken the assumption conditional constraint dependence in the absence of high-level cognitive information, improve the unity and integrity of salient target extraction, and strengthen the adaptability of the algorithm still needs further research and overcoming. technical difficulties.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a technical solution for a method for extracting the consistency of image salient target regions under natural background, which can make full use of the comprehensive differences between foreground and background in images, integrate the inherent relationship between salient targets, and weaken the influence of salience analysis process on the The dependence of the traditional assumption condition constraints, realizes the consistent extraction of multiple salient objects, guarantees the internal integrity of a single salient object, and the integrity of multi-objective extraction.

In order to achieve the above purpose, the technical solution provided by the present invention is an image salient target region extraction method based on iterative sparse representation, comprising the following steps:

Step 1, data preprocessing, set different SLIC superpixel numbers, perform multi-scale superpixel segmentation on the original image, use the saliency detection based on classical visual attention, and set the detection result as the initial saliency map SAL ₀ ;

Step 2, extract the salient features of the pixel-level original image, including color features, position features and gradient features;

Step 3, by calculating the mean value of all original pixel features in the superpixel region of each scale, obtain the superpixel region features under a single scale segmentation;

Step 4, for the segmentation result of a single scale, calculate the saliency map through recursive sparse representation, including the following sub-steps:

Step 4.1, superpixel initial saliency calculation, calculate the initial saliency level of each superpixel through mean calculation according to SAL ₀ ;

Step 4.2, extracting foreground samples, arranging the initial significance levels of superpixels in descending order, and taking the first p1% of superpixels as foreground samples D _f ;

Step 4.3, background sample extraction, arranging the initial significance levels of superpixels in ascending order, taking the first p2% superpixels as candidate background samples D _b1 , extracting superpixels touching the boundary of the image as candidate background samples D _b2 , background samples Calculated as follows:

D _b =D _b1 +D _b2 -D _f (1)

Step 4.4, double sparse representation and sparse residual calculation, use foreground samples and background samples as dictionaries to sparsely represent all superpixels and calculate the reconstruction residuals, the formula is as follows:

where i represents the superpixel number; F _i is the feature vector of the superpixel region; λ _b , λ _f are regular parameters; α _bi , α _fi are the foreground sparse representation result and the background sparse representation result respectively; ε _bi , ε _fi are respectively Foreground dilution reconstruction residual and background sparse reconstruction residual;

Step 4.5, saliency factor calculation, fuse ε _bi and ε _fi according to formula (6) and assign the superpixel fusion result to all original image pixels in it, and calculate the saliency factor map SAL _i ,

SAL _i =ε _bi /(ε _fi +σ ² ) (6)

where σ ² is a non-negative tuning parameter;

Step 4.6, recursive processing, calculate the correlation coefficient rela between the saliency factor map SAL _i and the initial saliency map SAL ₀ according to formula (7), if rela<K, then set SAL ₀ =SAL _i and repeat the whole process of step 4 ; If rela>K, the recursion ends, and the output current SAL _i is the saliency detection result at this scale; where K is the similarity judgment threshold,

rela=corr2(A,B) (7)

where corr2() is the correlation coefficient calculation function; A and B are the matrices or images to be compared; rela is the correlation coefficient between A and B, the larger the value, the more similar A and B, and vice versa, the greater the difference;

In step 5, the multi-scale saliency detection results are fused, and the saliency results under each single scale are combined linearly with equal weights to calculate the final saliency detection result.

Further, the salient features described in step 2 are RGB, Lab, x, y, and a total of 13-dimensional features of the first-order gradient and the second-order gradient, which are expressed as {R, G, B, L, a, b, x, y,f _x ,f _y ,f _xx ,f _yy ,f _xy }, where R, G, B, L, a, b have six-dimensional features as color features, R, G, B and L, a, b respectively are RGB and LAB color information; x, y are position information, x, y are the row and column coordinates of the pixel in the image; f _x , f _y , f _xx , f _yy , f _xy are gradient features, which respectively indicate that the pixel is in X, The first-order and second-order differences in the Y direction are calculated as follows:

f _x =(f(i+1,j)-f(i-1,j))/2

f _y =(f(i,j+1)-f(i,j-1))/2

f _xx = (f _x (i+1,j)-f _x (i-1,j))/2 (8)

f _yy =(f _y (i,j+1)-f _y (i,j-1))/2

f _xy = (f _x (i,j+1)-f _x (i,j-1))/2

Among them, f(i,j) is the image matrix, and i,j is the image pixel row and column number.

The present invention also correspondingly provides an image salient target region extraction system based on iterative sparse representation, comprising the following modules:

The preprocessing module is used for data preprocessing, setting different SLIC superpixel numbers, performing multi-scale superpixel segmentation on the original image, using saliency detection based on classical visual attention, and setting the detection result as the initial saliency map SAL ₀ ;

The saliency feature extraction module is used to extract the saliency features of pixel-level original images, including color features, position features and gradient features;

The superpixel region feature acquisition module is used to obtain the superpixel region features under a single scale segmentation by calculating the mean value of all original pixel features in the superpixel region of each scale;

The sparse representation module is used to calculate the saliency map through recursive sparse representation for the segmentation results of a single scale, including the following sub-modules:

The first submodule is used for the initial saliency calculation of superpixels, and calculates the initial saliency level of each superpixel through mean value calculation according to SAL ₀ ;

The second sub-module is used for foreground sample extraction, the initial significance levels of superpixels are sorted in descending order, and the first p1% superpixels are taken as foreground samples D _f ;

The third sub-module is used for background sample extraction, arranging the initial significance levels of superpixels in ascending order, taking the first p2% superpixels as candidate background samples D _b1 , and extracting superpixels touching the boundary of the image as candidate background samples D _b2 , the background sample calculation formula is as follows:

D _b =D _b1 +D _b2 -D _f (1)

The fourth sub-module is used for double sparse representation and sparse residual calculation. The foreground samples and background samples are used as dictionaries to sparsely represent all superpixels and calculate the reconstruction residual. The formula is as follows:

where i represents the superpixel number; F _i is the feature vector of the superpixel region; λ _b , λ _f are regular parameters; α _bi , α _fi are the foreground sparse representation result and the background sparse representation result respectively; ε _bi , ε _fi are respectively Foreground dilution reconstruction residual and background sparse reconstruction residual;

The fifth sub-module is used for saliency factor calculation. According to formula (6), ε _bi and ε _fi are fused, and the superpixel fusion result is assigned to all original image pixels in it, and the saliency factor map SAL _i is obtained by calculating,

SAL _i =ε _bi /(ε _fi +σ ² ) (6)

where σ ² is a non-negative tuning parameter;

The sixth sub-module, used for recursive processing, calculates the correlation coefficient rela between the saliency factor map SAL _i and the initial saliency map SAL ₀ according to formula (7). If rela<K, then make SAL ₀ =SAL _i and repeat the execution The whole process of step 4; if rela>K, the recursion ends, and the output current SAL _i is the saliency detection result under this scale; among them, K is the similarity judgment threshold,

rela=corr2(A,B) (7)

where corr2() is the correlation coefficient calculation function; A and B are the matrices or images to be compared; rela is the correlation coefficient between A and B, the larger the value, the more similar A and B, and vice versa, the greater the difference;

The detection result fusion module is used for the fusion of multi-scale saliency detection results, and performs equal-weight linear combination of saliency results under each single scale to calculate the final saliency detection result.

Further, the salient features described in the salient feature extraction module are RGB, Lab, x, y, and a total of 13-dimensional features of the first-order gradient and the second-order gradient, which are expressed as {R, G, B, L, a, b ,x,y,f _x ,f _y ,f _xx ,f _yy ,f _xy }, where R, G, B, L, a, b have six-dimensional features as color features, R, G, B and L, a , b are RGB and LAB color information respectively; x, y are position information, x, y are the row and column coordinates of pixels in the image; f _x , f _y , f _xx , f _yy , f _xy are gradient features, representing pixels respectively The first-order and second-order differences in the X and Y directions are calculated as follows:

f _x =(f(i+1,j)-f(i-1,j))/2

f _y =(f(i,j+1)-f(i,j-1))/2

f _xx = (f _x (i+1,j)-f _x (i-1,j))/2 (8)

f _yy =(f _y (i,j+1)-f _y (i,j-1))/2

f _xy = (f _x (i,j+1)-f _x (i,j-1))/2

Among them, f(i,j) is the image matrix, and i,j is the image pixel row and column number.

The method of the invention firstly uses multiple groups of SLIC segmentation methods with different pixel number parameters to perform superpixel segmentation on the original image, generates a group of segmented images with different sizes of superpixel regions, and establishes multi-scale source data. Then, for the segmentation results of each scale, the classical visual attention detection results are used as the initial saliency map to constrain the selection of foreground and background sample regions, and then the reconstruction residuals of each superpixel region are calculated as saliency factors through the sparse representation process. , and combined with recursive iterative operations to optimize the saliency detection result map at a single scale, and finally obtain the final salient target and detection result through multi-scale saliency map fusion. The technical scheme of the present invention has the following advantages:

1) The image is divided into superpixel images of multiple scales by multiple groups of SLIC segmenters. On the one hand, the SLIC method can effectively retain the image contour information, which helps the saliency detection process to maintain the consistency within the same target area; Multi-scale segmentation enables the algorithm to have better adaptability and robustness to the detection of objects of different sizes.

2) Calculate the pixel (regional) saliency through the double sparse representation process based on the foreground dictionary and the background dictionary. On the one hand, the residual of the reconstruction process is used as the saliency level indicator to determine the visual saliency similarity level between pixels from a global perspective. Different from the traditional methods based on contrast and image boundary constraints, it can effectively improve the problem of incomplete target detection; in addition, the calculation process of double sparse representation can also analyze the attributes of each pixel more comprehensively to determine its significance level. Further improve the robustness of the algorithm.

3) Through the recursive optimization process, the dependence of the algorithm on the initial saliency map generated based on the classical visual attention model can be weakened to a certain extent, which is helpful to improve the reliability of the algorithm.

Description of drawings

Fig. 1 is a comparison of the detection results of the salient target in the image processed by the method of the present invention and the traditional method in the embodiment of the present invention, (a) is the input image, (b) is the true value of the salient target, (c) is the detection result of the traditional method based on local contrast , (d) is the detection result based on the global contrast method, (e) is the detection result of the image boundary constraint method, and (f) is the detection result of the method of the present invention.

FIG. 2 is a flowchart of an embodiment of the present invention.

Detailed ways

The specific technical solutions of the present invention will be described below according to the accompanying drawings and embodiments.

The invention proposes an image salient target region extraction method based on iterative sparse representation. This method extracts the target area that can most attract human visual attention by analyzing the saliency of the image, which can play the role of effective data optimization and data compression, and is the basic link of many image processing problems. The study found that traditional image salient object extraction methods based on local contrast, global contrast and image boundary constraints usually have strong dependence on the corresponding constraints, and are prone to inconsistent internal saliency of single target and incomplete multi-target detection. and the problem that the extraction of salient objects with image boundaries is not clear, as shown in Figure 1(c)-(e). This method uses double sparse representation and reconstruction residuals to calculate the pixel saliency level, optimizes the detection results by means of a recursive iterative process, and improves the applicability of the algorithm by fusing the multi-scale detection results. As shown in Figure 1(f), the method of the present invention The single-target saliency in the detection results has better consistency, and is more complete for multi-target detection. At the same time, it can also improve the traditional method to a certain extent. The lack of missed detection of salient targets close to the image boundary is fully confirmed by the embodiment. It has better detection performance than traditional general salient target extraction methods. As shown in Figure 2, the specific implementation method provided by the embodiment includes the following steps:

Step 1, data acquisition. Download the saliency detection open source dataset raw images and saliency object ground-truth data.

Step 2, data preprocessing. Multi-scale segmentation is performed on the original image, and a classical visual attention model is used for saliency detection to generate an initial saliency map.

Step 3: Extract the 13-dimensional features of the original image pixels, namely {R, G, B, L, a, b, x, y, f _x , f _y , f _xx , f _yy , f _xy }.

Step 4: Extract the superpixel region features under single-scale segmentation through mean calculation, that is, F={mR,mG,mB,mL,ma,mb,mx,my,mf _x ,mf _y ,mf _xx ,mf _yy , mf _xy }, where F is the region feature vector, mX (X=R, G, B, L, a, b, x, y, f _x , f _y , f _xx , f _yy , f _xy ) is the superpixel region The mean of the X attribute of all the original pixels in the image.

Step 5, for the segmentation result of a single scale, calculate the saliency map through recursive sparse representation, including the following sub-steps:

Step 5.1, obtain the initial saliency map of superpixels through mean calculation.

Step 5.2, extract some superpixel regions with high initial saliency level as foreground samples. The initial significance levels of the superpixels are arranged in descending order, and the first p1% of the superpixels are taken as the foreground samples D _f . In this embodiment, p1 is taken as 20, and those skilled in the art can select an appropriate value as needed;

Step 5.3, combine the initial saliency map and image boundary constraints to extract background samples. The initial significance levels of the superpixels are arranged in ascending order, and the first p2% of the superpixels are taken as the candidate background samples D _b1 . The superpixel is used as the candidate background sample D _b2 , and the calculation formula of the background sample is as follows:

D _b =D _b1 +D _b2 -D _f (1)

Step 5.4: Use foreground samples and background samples to perform two sets of sparse representations on all superpixels in the image, and calculate the corresponding reconstruction residuals. The formula is as follows:

where i represents the number of the superpixel; F _i is the feature vector of the superpixel region; λ _b , λ _f are regular parameters, in this embodiment λ _b , λ _f both take 0.01; α _bi , α _fi are the foreground sparse representation results, respectively and background sparse representation results; ε _bi , ε _fi are the foreground dilution reconstruction residual and the background sparse reconstruction residual, respectively;

Step 5.5, fuse the two groups of reconstruction residuals to generate superpixel saliency factors, and obtain the original image saliency factor map based on the requirement that the saliency of the original image pixels in the superpixel area is consistent. According to formula (6), ε _bi and ε _fi are fused and the superpixel fusion result is assigned to all the original image pixels in it, and the saliency factor map SAL _i is obtained by calculating,

SAL _i =ε _bi /(ε _fi +σ ² ) (6)

Wherein σ ² is a non-negative adjustment parameter, which is taken as 0.1 in this embodiment, and those skilled in the art can select an appropriate value as required;

Step 5.6, compare the saliency factor map obtained in 5.5 with the initial saliency map in 5.1, execute the recursive processing process, and output the saliency detection result at the current scale at the end of the recursion. Calculate the correlation coefficient rela between the saliency factor map SAL _i and the initial saliency map SAL ₀ according to formula (7). If rela<K, then set SAL ₀ =SAL _i and repeat the whole process of step 4; if rela>K, Then the recursion ends, and the output current SAL _i is the significance detection result under this scale; wherein, K is the similarity judgment threshold, and in this embodiment, K takes 0.99, and those skilled in the art can select an appropriate value as needed;

rela=corr2(A,B) (7)

where corr2() is the correlation coefficient calculation function; A and B are the matrices or images to be compared; rela is the correlation coefficient between A and B, the larger the value, the more similar A and B, and vice versa, the greater the difference;

Step 6: Perform mean fusion on the multi-scale saliency detection results to generate the final target extraction result.

Theoretically, in the implementation of the entire technical solution of the present invention, the overall consistency extraction of the salient target area in the natural background image is realized under the support of the principle of sparse representation. Different from the traditional salient target detection method based on contrast and image boundary constraints, the present invention makes full use of the comprehensive difference between the foreground and background in the image, integrates the inherent relationship within the salient target and between multiple salient targets, and tries to avoid the traditional contrast-based method. Constraints and image boundary constraint detection methods face the problem of hypothetical conditional dependence. The sparse representation is used as the image pixel consistency analysis method, and the sparse reconstruction residual is used as the pixel difference index. The sparse reconstruction based on the image foreground and background dictionaries is used. The residual is used as a saliency factor to achieve consistent extraction of multiple salient targets, ensuring the integrity of a single salient target and the integrity of multi-target extraction.

During specific implementation, the technical solution of the present invention can realize the automatic running process based on computer software technology, and can also realize the corresponding system in a modular way. An embodiment of the present invention provides an image salient target region extraction system based on iterative sparse representation, including the following modules:

The preprocessing module is used for data preprocessing, setting different SLIC superpixel numbers, performing multi-scale superpixel segmentation on the original image, using saliency detection based on classical visual attention, and setting the detection result as the initial saliency map SAL ₀ ;

The saliency feature extraction module is used to extract the saliency features of pixel-level original images, including color features, position features and gradient features;

The superpixel region feature acquisition module is used to obtain the superpixel region features under a single scale segmentation by calculating the mean value of all original pixel features in the superpixel region of each scale;

The sparse representation module is used to calculate the saliency map through recursive sparse representation for the segmentation results of a single scale, including the following sub-modules:

The first submodule is used for the initial saliency calculation of superpixels, and calculates the initial saliency level of each superpixel through mean value calculation according to SAL ₀ ;

The second sub-module is used for foreground sample extraction, the initial significance levels of superpixels are sorted in descending order, and the first p1% superpixels are taken as foreground samples D _f ;

The third sub-module is used for background sample extraction, arranging the initial significance levels of superpixels in ascending order, taking the first p2% superpixels as candidate background samples D _b1 , and extracting superpixels touching the boundary of the image as candidate background samples D _b2 , the background sample calculation formula is as follows:

D _b =D _b1 +D _b2 -D _f (1)

The fourth sub-module is used for double sparse representation and sparse residual calculation. The foreground samples and background samples are used as dictionaries to sparsely represent all superpixels and calculate the reconstruction residual. The formula is as follows:

where i represents the superpixel number; F _i is the feature vector of the superpixel region; λ _b , λ _f are regular parameters; α _bi , α _fi are the foreground sparse representation result and the background sparse representation result respectively; ε _bi , ε _fi are respectively Foreground dilution reconstruction residual and background sparse reconstruction residual;

The fifth sub-module is used for saliency factor calculation. According to formula (6), ε _bi and ε _fi are fused, and the superpixel fusion result is assigned to all original image pixels in it, and the saliency factor map SAL _i is obtained by calculating,

SAL _i =ε _bi /(ε _fi +σ ² ) (6)

where σ ² is a non-negative tuning parameter;

The sixth sub-module, used for recursive processing, calculates the correlation coefficient rela between the saliency factor map SAL _i and the initial saliency map SAL ₀ according to formula (7). If rela<K, then set SAL ₀ =SAL _i and repeat the execution The whole process of step 4; if rela>K, the recursion ends, and the output current SAL _i is the saliency detection result under this scale; among them, K is the similarity judgment threshold,

rela=corr2(A,B) (7)

where corr2() is the correlation coefficient calculation function; A and B are the matrices or images to be compared; rela is the correlation coefficient between A and B, the larger the value, the more similar A and B, and vice versa, the greater the difference;

The detection result fusion module is used for the fusion of multi-scale saliency detection results, and performs equal-weight linear combination of saliency results under each single scale to calculate the final saliency detection result.

The saliency features described in the saliency feature extraction module are RGB, Lab, x, y, and a total of 13-dimensional features of the first-order gradient and the second-order gradient, which are expressed as F={R,G,B,L,a,b, x, y, f _x , f _y , f _xx , f _yy , f _xy }, where R, G, B, L, a, b have six-dimensional features as color features, R, G, B and L, a, b are RGB and LAB color information respectively; x, y are position information, x, y are the row and column coordinates of the pixel in the image; f _x , f _y , f _xx , f _yy , f _xy are gradient features, respectively indicating that the pixel is in the image The first-order and second-order differences in the X and Y directions are calculated as follows:

f _x =(f(i+1,j)-f(i-1,j))/2

f _y =(f(i,j+1)-f(i,j-1))/2

f _xx = (f _x (i+1,j)-f _x (i-1,j))/2 (8)

f _yy =(f _y (i,j+1)-f _y (i,j-1))/2

f _xy = (f _x (i,j+1)-f _x (i,j-1))/2

Among them, f(i,j) is the image matrix, and i,j is the image pixel row and column number.

The specific implementation of each module can refer to the corresponding steps, which will not be described in the present invention.

The foregoing description of the embodiments only illustrates the basic technical solutions of the present invention, and is not limited to the foregoing embodiments. A person or team skilled in the art to which the present invention pertains can make any simple modifications, additions, equivalent changes or modifications to the described specific embodiments without departing from the basic spirit of the present invention or beyond the scope defined by the claims.

Claims (4)

1. The image salient object region extraction method based on the iterative sparse representation is characterized by comprising the following steps of:

step 1, preprocessing data, setting different SLIC superpixel numbers, carrying out multi-scale superpixel segmentation on an original image, using saliency detection based on classical visual attention, and setting a detection result as an initial saliency map SAL₀；

Step 2, extracting the salient features of the pixel-level original image, wherein the salient features comprise color features, position features and gradient features;

step 3, calculating the average value of all original pixel features in the superpixel region of each scale to obtain the superpixel region features under single-scale segmentation;

step 4, aiming at the segmentation result of a single scale, calculating a saliency map through recursive sparse representation, and comprising the following substeps:

step 4.1, superpixel initial saliency calculation according to SAL₀Solving the initial significance level of each super pixel through mean value calculation;

step 4.2, extracting foreground samples, performing descending order arrangement on the initial significance level of the superpixels, and taking the front p 1% superpixels as foreground samples D_f；

Step 4.3, extracting a background sample, arranging the initial significance levels of the superpixels in an ascending order, and taking the front p2% superpixels as an alternative background sample D_b1Extracting superpixels contacting the image boundary as alternative background samples D_b2The background sample calculation formula is as follows:

D_b＝D_b1+D_b2-D_f (1)

step 4.4, performing double sparse representation and sparse residual calculation, wherein the foreground sample and the background sample are respectively used as dictionaries to perform sparse representation on all the superpixels, and a reconstructed residual is calculated, wherein the formula is as follows:

wherein i represents a super pixel number; f_iIs a feature vector of the superpixel region; lambda [ alpha ]_b，λ_fIs a regularization parameter; alpha is alpha_bi，α_fiRespectively representing a foreground sparse representation result and a background sparse representation result; epsilon_bi，ε_fiRespectively a foreground dilution reconstruction residual error and a background sparse reconstruction residual error;

step 4.5, calculating the significance factor, and aligning epsilon according to a formula (6)_biAnd ε_fiFusing, giving the super-pixel fusion result to all original image pixels in the super-pixel fusion result, and calculating to obtain a significant factor graph SAL_i，

SAL_i＝ε_bi/(ε_fi+σ²) (6)

Wherein sigma²A non-negative tuning parameter;

step 4.6, recursive processing is carried out, and the significance factor graph SAL is calculated according to a formula (7)_iAnd an initial saliency map SAL₀Relative coefficient between them, if rela is less than K, SAL is ordered₀＝SAL_iAnd the whole process of the step 4 is repeatedly executed; if rela > K, the recursion is ended, and the current SAL is output_iThe significance detection result at the scale is obtained; wherein K is a similarity determination threshold value,

rela＝corr2(A，B) (7)

wherein corr2() is the correlation coefficient calculation function; a and B are matrixes or images to be compared; rela is a correlation coefficient between A and B, the larger the value is, the more similar A and B are, otherwise, the difference is larger;

and 5, fusing multi-scale significance detection results, performing equal-weight linear combination on significance results under each single scale, and calculating a final significance detection result.

2. The image salient object region extraction method based on iterative sparse representation as claimed in claim 1, wherein the salient features in step 2 are RGB, Lab, x, y, and 13-dimensional features with first-order gradient and second-order gradient in total, and are represented as { R, G, B, L, a, B, x, y, f_x，f_y，f_xx，f_yy，f_xyThe color feature is a six-dimensional feature of R, G, B, L, a and B, and the R, G, B and L, a and B are RGB and Lab color information respectively; x and y are position information, and x and y are row-column coordinates of pixels in the image; f. of_x，f_yf_xx，f_yy，f_xFor the gradient feature, the first and second order differences of the pixel in the X and Y directions are respectively expressed, and the calculation formula is as follows:

where f (i, j) is the image matrix and i, j is the image pixel row column number.

3. The image salient object region extraction system based on the iterative sparse representation is characterized by comprising the following modules:

a preprocessing module for preprocessing data, setting different SLIC superpixel numbers, performing multi-scale superpixel segmentation on the original image, using the saliency detection based on classical visual attention, setting the detection result as an initial saliency map SAL₀；

The salient feature extraction module is used for extracting salient features of the pixel-level original image, wherein the salient features comprise color features, position features and gradient features;

the super-pixel region feature acquisition module is used for calculating the mean value of all original pixel features in the super-pixel region of each scale to obtain the super-pixel region features under single-scale segmentation;

the sparse representation module is used for calculating the saliency map through recursive sparse representation aiming at the segmentation result of a single scale, and comprises the following sub-modules:

a first sub-module for superpixel initial saliency calculation according to SAL₀Solving the initial significance level of each super pixel through mean value calculation;

the second sub-module is used for extracting the foreground sample, performing descending order arrangement on the initial significance level of the superpixels, and taking the front p 1% superpixels as the foreground sample D_f；

The third sub-module is used for extracting a background sample, the initial significance levels of the super pixels are arranged in an ascending order, and the front p 20% of super pixels are taken as an alternative background sample D_b1Extracting superpixels contacting the image boundary as alternative background samples D_b2The background sample calculation formula is as follows:

D_b＝D_b1+D_b2-D_f (1)

the fourth submodule is used for dual sparse representation and sparse residual calculation, all the superpixels are sparsely represented and reconstructed residual is calculated by taking the foreground sample and the background sample as dictionaries, and the formula is as follows:

wherein i represents a super pixel number; f_iIs a feature vector of the superpixel region; lambda [ alpha ]_b，λ_fIs a regularization parameter; alpha is alpha_bi，α_fiRespectively representing a foreground sparse representation result and a background sparse representation result; epsilon_bi，ε_fiRespectively a foreground dilution reconstruction residual error and a background sparse reconstruction residual error;

a fifth submodule for calculating the significance factor, for ε according to equation (6)_biAnd ε_fiFusing, giving the super-pixel fusion result to all original image pixels in the super-pixel fusion result, and calculating to obtain a significant factor graph SAL_i，

SAL_i＝ε_bi/(ε_fi+σ²) (6)

Wherein sigma²A non-negative tuning parameter;

a sixth sub-module for recursive processing for calculating the significance factor graph SAL according to equation (7)_iAnd an initial saliency map SAL₀Relative coefficient between them, if rela is less than K, SAL is ordered₀＝SAL_iAnd the whole process of the step 4 is repeatedly executed; if rela > K, the recursion is ended, and the current SAL is output_iThe significance detection result at the scale is obtained; wherein K is a similarity determination threshold value,

rela＝corr2(A，B) (7)

wherein corr2() is the correlation coefficient calculation function; a and B are matrixes or images to be compared; rela is a correlation coefficient between A and B, the larger the value is, the more similar A and B are, otherwise, the difference is larger;

and the detection result fusion module is used for fusing multi-scale significance detection results, performing equal-weight linear combination on significance results under each single scale, and calculating a final significance detection result.

4. The iterative sparse representation-based image salient object region extraction system according to claim 3, wherein the salient features in the salient feature extraction module are RGB, Lab, x, y and 13-dimensional features with first-order gradient and second-order gradient in total, and are represented as { R, G, B, L, a, B, x, y, f_x，f_y，f_xx，f_yy，f_xyThe color feature is a six-dimensional feature of R, G, B, L, a and B, and the R, G, B and L, a and B are RGB and Lab color information respectively; x and y are position information, and x and y are row-column coordinates of pixels in the image; f. of_x，f_y，f_xx，f_yy，f_xyFor the gradient feature, the first and second order differences of the pixel in the X and Y directions are respectively expressed, and the calculation formula is as follows:

where f (i, j) is the image matrix and i, j is the image pixel row column number.

Images