CN105069774B - The Target Segmentation method of optimization is cut based on multi-instance learning and figure - Google Patents

Abstract

The invention discloses a target segmentation method based on multi-instance learning and graph-cut optimization: Step 1: use the multi-instance learning method to carry out saliency model modeling on the training image, and use the saliency model to carry out saliency model modeling on the packages and examples in the test image Predict, get the saliency detection result of the test image; Step 2: Introduce the saliency detection result of the test image into the graph cut framework, optimize the graph cut framework according to the example feature vector and the label of the example package, and solve the suboptimal graph cut optimization solution to obtain an accurate segmentation of the target. The present invention adopts the method of multi-instance learning to establish a saliency detection model to make it suitable for specific types of images, and uses the results of saliency detection in an image segmentation method based on graph theory to guide image segmentation, and the frame link of the graph cut model It is optimized and solved by agglomerative hierarchical clustering algorithm, so that the segmentation results can better conform to the output of semantic awareness, and obtain accurate target segmentation results.

Description

Object Segmentation Method Based on Multiple Instance Learning and Graph Cut Optimization

technical field

The invention belongs to the field of image processing, and relates to an image segmentation method, in particular to an object segmentation method based on multi-instance learning and graph cut optimization.

Background technique

Image target segmentation is an important research direction in the field of computer vision, and it is also an important basis for applications such as visual inspection, tracking and recognition. The quality of its segmentation affects the performance of the entire visual system to a large extent. However, due to the lack of deep understanding of the human visual system, image segmentation has also become a classic problem in the field of computer vision. The human visual system is capable of selectively paying attention to the primary content of the observed scene while ignoring other secondary content. This selective attention mechanism of vision makes efficient information processing possible, and also inspires computer vision researchers to find another way from the perspective of attention mechanism, so the image segmentation model with human visual characteristics will become a new field of image segmentation. Research hotspots.

Saliency detection from the perspective of computer vision is mainly divided into bottom-up and top-down methods. At present, most saliency detection is based on unsupervised models. There are problems such as the lack of learning ability of the defined model, the calculation of saliency can not reflect the visual attention mechanism well, and the lack of adaptability and poor robustness to specific types of images. ; while the single cost function-based graph cut algorithm for object segmentation also has problems such as high computational complexity, low segmentation efficiency and local segmentation accuracy.

Contents of the invention

In view of the deficiencies in the above-mentioned prior art, the object of the present invention is to propose a target segmentation method based on multi-instance learning and graph cut optimization, and adopt the method of multi-instance learning to establish a saliency detection model to make it suitable for specific types of objects. image, and use the results of saliency detection in the image segmentation method based on graph theory to guide image segmentation, optimize the graph cut model framework and many other links, and use the agglomerative hierarchical clustering algorithm as the solution method for graph cut optimization , so that the segmentation results can better conform to the output of semantic awareness, and obtain accurate target segmentation results.

In order to achieve the above object, the present invention adopts the following technical solutions to solve it:

A target segmentation method based on multi-instance learning and graph cut optimization, including the following steps:

Step 1: Use the method of multi-instance learning to model the saliency model on the training image, and use the saliency model to predict the packages and examples in the test image, and obtain the saliency detection result of the test image; specifically include:

Step 11, preprocessing the training image, and extracting image brightness gradient features and color gradient features;

Step 12, introducing multi-instance learning into the image saliency detection to obtain the saliency detection result of the test image;

Step 2: Introduce the saliency detection results of the test image into the graph cut framework, optimize the graph cut framework according to the example feature vector and the label of the example package, solve the suboptimal solution of the graph cut optimization, and obtain the accurate segmentation of the target.

Further, in the step 11, the training image is preprocessed, and the brightness gradient feature and the color gradient feature are extracted, specifically including:

Step 111, performing color space conversion and quantization preprocessing of each component on the training image to obtain normalized brightness component L and color components a, b;

Step 112, calculating the brightness gradient of each pixel in the matrix of the brightness component L;

Step 113, calculating the color gradient of each pixel in the matrix of color component a and color component b respectively.

3. The object segmentation method based on multi-instance learning and graph cut optimization according to claim 2, wherein the step 111 is specifically as follows:

First, gamma correction is performed on the training image to achieve nonlinear adjustment of the image color components, and the training image is converted from the RGB color space to the Lab color space; then the brightness component L of the training image in the Lab color space and the two The color components a and b are normalized to obtain the normalized brightness component L and the color components a and b.

Further, the step 112 specifically includes steps A-D:

A. Construct the weight matrix Wights<> of 3 scales;

B. Construct the index map matrix Slice_map<> of 3 scales; the index map matrix Slice_map<> of each scale Slice_map<> has the same dimension as the weight matrix Wights<> of the corresponding scale; select 8 directions (0 °, 22.5 °, 45 °, 67.5°, 90°, 112.5°, 135°, 157.5°) divide the matrix into 16 areas, and the value of the elements in each area is the same as the number 0~15 of the area;

C. Multiply each index map matrix Slice_map<> with the elements in the weight matrix Wights<> of the corresponding scale one by one to obtain the matrix of the corresponding scale, that is, the neighborhood gradient operator;

D. Using the neighborhood gradient operator, calculate the brightness gradient of a pixel to be calculated in the matrix of the brightness component L;

Further, the step A is specifically as follows:

Construct the weight matrix Wights<> of three scales respectively; the weight matrix Wights<> is a square matrix with the number of rows and columns equal to 2r+1; the elements in the weight matrix Wights<> are either 0 or 1 , the elements equal to 1 are distributed within the range of the disk with the central element (r+1, r+1) as the center of the square matrix and the radius r as the circle, forming the inscribed circle of the square matrix, and the rest of the elements in the square matrix are 0 ; The 3 scales are r=3, r=5 and r=10 respectively.

Further, the step D is specifically as follows:

① For a certain scale, take a pixel to be sought in the matrix of the luminance component L obtained in step 111 as the center, and perform a calculation with each luminance component within the neighborhood of the pixel to be sought by a neighborhood gradient operator of a certain scale Dot multiplication to obtain the matrix Neibor<> within the neighborhood of the pixel to be sought; select a straight line in the vertical direction (90°) as the dividing line, and divide the disk in the neighborhood gradient operator into a left semicircle and a right semicircle, The left semicircle includes the 0th sector to the 7th sector, and the right semicircle includes the 8th sector to the 15th sector; the elements of the matrix Neibor<> corresponding to each semicircle form a histogram and normalize it, respectively Recorded as Slice_hist ₁ <> and Slice_hist ₂ <>; H ₁ represents the histogram corresponding to the left semicircle area, H ₂ represents the histogram corresponding to the right semicircle area, i is the value of the bin of the histogram, defined as [0 ,24], that is, the brightness range;

②Calculate the difference between the two normalized histograms through the chi-square distance shown in formula (1), that is, obtain the brightness gradient in the vertical direction of a pixel to be obtained at a certain scale;

After calculating the luminance gradient in the vertical direction of a certain scale, select the straight lines in other directions as the dividing line to obtain the luminance gradient in all other directions of a certain scale of the pixel to be obtained; then calculate in the same way as step D Obtain the brightness gradient of the pixel to be requested in all directions on other scales; after completing the calculation of the brightness gradient of the pixel to be requested in all directions on all scales, the final brightness gradient of the pixel to be requested is calculated by formula (2) :

f(x,y,r,n_ori; r=3,5,10; n_ori=1,2,...8)->Brightness Gradient(x,y) (2)

In the formula, f is a mapping function, (x, y) is any pixel point to be obtained, r represents the selected scale, n_ori represents the selected direction; Brightness Gradient(x, y) is the pixel point (x, y) The final brightness gradient; the corresponding rule of f is to select the maximum brightness gradient value in each direction in 3 scales as the brightness gradient value in this direction, and sum the brightness gradients in 8 directions to obtain the pixel point (x, y) The final brightness gradient of .

Further, in the step 113, the calculation of the color gradient is similar to the calculation of the brightness gradient, the difference is that the color gradient feature is the color gradient a and b for the two color components; the three selected scales are r=5, r=10 and r=20; therefore, the sizes of the corresponding weight matrix and map index matrix are 11*11, 21*21 and 41*41 respectively; the calculation of the color gradient of the two color components and the brightness gradient adopt the same The calculation method obtains the final color gradient of each pixel to be calculated in the matrix of color components a and b.

Further, in the step 12, the multi-instance learning is introduced into the image saliency detection to obtain the saliency detection result of the test image, which specifically includes steps 121 and 122:

Step 121, using the brightness and color gradient features obtained by the method described in step 11, combined with the multi-instance learning EMDD algorithm to realize the learning of the training set, and obtain a learned saliency detection model;

Step 122, substituting the test image into the learned saliency detection model to obtain a saliency detection result of the test image.

Further, said step 2 specifically includes the following steps:

Step 21, the saliency detection result of the image obtained in step 1 is used as the input of the graph cut algorithm, and the weight function shown in formula (3) is constructed according to the saliency mark of the package and the example feature vector; and the formula (4) is obtained The optimized graph cut cost function shown;

In formula (3), w _ij represents the similarity of visual features of the regions corresponding to sample package i and package j, Salien(i) and Salien(j) respectively represent the normalized saliency values of region i and region j, σ In order to adjust the sensitive parameter of visual feature difference, the value is 10-20; the similarity weight between region i and itself is 0; the similarity matrix W={w _ij } is a symmetric matrix whose diagonal is 0, and w _ij ∈ [0,1]; f _i , f _j represent the example feature vectors corresponding to i and j example packages respectively, that is, the brightness gradient feature and the color gradient feature vector of the image are synthesized into a 3-dimensional combination vector Mixvector _i ={BrightnessGradient _i , ColorGradient _i }, then Sim(f _i , f _j )=‖Mixvector _i -Mixvector _j ‖ ₂ . In the graph cut framework represented by formula (4), D is an N-dimensional diagonal matrix, and the elements on the diagonal U={U ₁ , U ₂ ,...,U _i ,...,U _j ,...U _N } is the segmentation state vector, and each vector component U _i represents the segmentation state of area i; formula (4 ) represents the visual similarity between region i and region j, and the denominator represents the visual similarity within region i;

Step 22, solving the segmentation state vector corresponding to the minimum eigenvalue of R(U), that is, obtaining the optimal segmentation result of the image.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

Figure 2 is a comparison of the results of the test image segmentation by multiple methods. Among them, the subgraphs (a-1) to (a-4) are the original test images, the subgraphs (b-1) to (b-4) are the segmentation results of the spectral segmentation algorithm based on multi-scale graph decomposition, and the subgraphs (c‐1) to (c‐4) are the segmentation results directly using the agglomerative hierarchical clustering algorithm, and the subgraphs (d‐1) to (d‐4) are the segmentation results of the method of the present invention.

Fig. 3 is a schematic diagram of the left and right partitions of the disk.

Fig. 4 is a schematic diagram of the histogram of H ₁ and H ₂ .

Fig. 5 is a schematic diagram of changing the direction of the boundary line of the disk.

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

detailed description

As shown in Figure 1, the object segmentation method based on multi-instance learning and graph cut optimization of the present invention specifically includes the following steps:

Step 1: Use the method of multi-instance learning to model the saliency model on the training image, and use the saliency model to predict the packages and examples in the test image, and obtain the saliency detection result of the test image;

Step 2: Introduce the saliency of the test image into the graph cut framework, optimize the graph cut framework according to the example feature vector and the label of the example package, use the agglomerative hierarchical clustering algorithm to solve the suboptimal solution of the graph cut optimization, and obtain the accurate segmentation of the target .

Further, said step 1 includes step 11 and step 12:

Step 11, preprocessing the training image, and extracting image brightness gradient features and color gradient features;

Step 12, introduce multi-instance learning into image saliency detection, and obtain the saliency detection result of the test image.

Further, in the step 11, the training image is preprocessed, and the brightness gradient feature and the color gradient feature are extracted, specifically including steps 111 to 113:

Step 111, carry out the conversion of color space and the quantitative preprocessing of each component to the training image, obtain the normalized brightness component L and color components a, b; specifically as follows:

First, gamma correction is performed on the training image to achieve nonlinear adjustment of the image color components, and the training image is converted from the RGB color space to the Lab color space; then the brightness component L of the training image in the Lab color space and the two The color components a and b are normalized to obtain the normalized brightness component L and the color components a and b;

Step 112, calculating the brightness gradient of each pixel in the matrix of the brightness component L. Specifically include steps A-D:

A. Construct the weight matrix Wights<> of 3 scales. details as follows:

Construct the weight matrix Wights<> of three scales respectively; the weight matrix Wights<> is a square matrix with the number of rows and columns equal to 2r+1; the elements in the weight matrix Wights<> are either 0 or 1 , the elements equal to 1 are distributed within the range of the disk with the central element (r+1, r+1) as the center of the square matrix and the radius r as the circle, forming the inscribed circle of the square matrix, and the rest of the elements in the square matrix are 0 ; In the present invention, when the three scales are respectively r=3, r=5 and r=10, the corresponding weight matrix Wights<> is as follows:

B. Construct the index map matrix Slice_map<> of three scales; the index map matrix Slice_map<> of each scale and the weight matrix Wights<> of the corresponding scale have the same dimension, that is, each index map Slice_map<> matrix is also the number of rows and a square matrix whose number of columns is 2r+1; choose 8 directions (0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, 157.5°) to divide the matrix into 16 regions, each The values of the elements in a region are the same as the numbers 0-15 of the region; the purpose of establishing the index map matrix Slice_map<> is to realize the fast positioning of the partition. In the present invention, the three index map matrices Slice_map<> are as follows:

C. Multiply each index map matrix Slice_map<> with the elements in the weight matrix Wights<> of the corresponding scale one by one to obtain a matrix of the corresponding scale, that is, the neighborhood gradient operator. The neighborhood gradient operators at the three scales are as follows:

D. Using a neighborhood gradient operator to calculate the brightness gradient of a pixel to be calculated in the matrix of the brightness component L. details as follows:

① For a certain scale, take a pixel to be sought in the matrix of the luminance component L obtained in step 111 as the center, and perform a calculation with each luminance component within the neighborhood of the pixel to be sought by a neighborhood gradient operator of a certain scale Dot multiplication to obtain the matrix Neibor<> within the neighborhood of the pixel to be sought; select a straight line in the vertical direction (90°) as the dividing line, and divide the disk in the neighborhood gradient operator into a left semicircle and a right semicircle, The left semicircle includes the 0th sector to the 7th sector, and the right semicircle includes the 8th sector to the 15th sector; the elements of the matrix Neibor<> corresponding to each semicircle form a histogram and normalize it, respectively Recorded as Slice_hist ₁ <> and Slice_hist ₂ <>; as shown in Figure 4. H ₁ represents the histogram corresponding to the left semicircle area, H ₂ represents the histogram corresponding to the right semicircle area, and i is the value of the bin of the histogram, which is defined as [0,24], that is, the brightness range.

②Calculate the difference between the two normalized histograms through the chi-square distance shown in formula (1), that is, obtain the brightness gradient in the vertical direction of a pixel to be obtained at a certain scale;

After calculating the brightness gradient in the vertical direction of a certain scale, as shown in Figure 5, select the straight lines in other directions as the dividing line to obtain the brightness gradient in all other directions of a certain scale of the pixel to be obtained; then according to In step D, the brightness gradients in all directions on other scales of the pixel to be obtained are calculated in the same manner. After completing the calculation of the brightness gradient of the pixel to be requested in all scales and directions, the final brightness gradient of the pixel to be requested is calculated by formula (2):

f(x,y,r,n_ori; r=3,5,10; n_ori=1,2,...8)->Brightness Gradient(x,y) (2)

In the formula, f is a mapping function, (x, y) is any pixel point to be obtained, r represents the selected scale, n_ori represents the selected direction; Brightness Gradient(x, y) is the pixel point (x, y) The final brightness gradient; the corresponding rule of f is to select the maximum brightness gradient value in each direction in 3 scales as the brightness gradient value in this direction, and sum the brightness gradients in 8 directions to obtain the pixel point (x, y) The final brightness gradient of ;

Step 113, calculating the color gradient of each pixel in the matrix of color component a and color component b respectively. details as follows:

The calculation of the color gradient is similar to the calculation of the brightness gradient, the difference is that the color gradient feature is the color gradient for two color components, that is, the color components a and b in the Lab color space; the difference from the calculation of the brightness gradient is that the selection The three scales of are r=5, r=10 and r=20 respectively; therefore, the sizes of the corresponding weight matrix and map index matrix are 11*11, 21*21 and 41*41 respectively; the two color components The same calculation method is used for the calculation of the color gradient and the brightness gradient, and the final color gradient of each pixel to be obtained in the color component a and b matrix is obtained.

Further, in step 12, the multi-instance learning is introduced into the image saliency detection to obtain the saliency detection result of the test image, specifically including steps 121 and 122:

Step 121, using the brightness and color gradient features obtained by the method described in step 11, combined with the multi-instance learning EMDD algorithm to learn the training set, and obtain a learned saliency detection model. Specific steps are as follows:

First, the super-segmentation method is used to segment the training image, so that the minimum number of pixels contained in each area is 200; each area is regarded as a bag, and each area is randomly sampled, and the pixels in the sampled area are regarded as As an example, extract the corresponding brightness gradient feature and color gradient feature vector as the sampling example feature vector; according to the sampling example feature vector, use the multi-example learning method EMDD algorithm to train the classifier, and obtain a learned saliency detection model;

Step 122, substituting the test image into the learned saliency detection model to obtain a saliency detection result of the test image.

For each test image, use the same process as step 11 to preprocess the test image to obtain the brightness gradient feature and color gradient feature; then use the super-segmentation method to segment the test image, so that the minimum pixel contained in each area The number is 200; treat each area as a package and randomly sample each area, the pixels in the sampled area are taken as examples, and the corresponding brightness gradient features and color gradient feature vectors are extracted as sampling example feature vectors, using The learned saliency detection model obtained in step 121 obtains the saliency example feature vector and the saliency of each package, so as to obtain the saliency detection result of the test image.

Further, the step 2 specifically includes the following steps:

Step 21, the saliency detection result of the image obtained in step 1 is used as the input of the graph cut algorithm, and the weight function shown in formula (3) is constructed according to the saliency mark of the package and the example feature vector; and the formula (4) is obtained The optimized graph cut cost function shown;

In formula (3), w _ij represents the similarity of visual features of the regions corresponding to sample package i and package j, Salien(i) and Salien(j) respectively represent the normalized saliency values of region i and region j, σ In order to adjust the sensitive parameter of visual feature difference, the value is 10-20; the similarity weight between region i and itself is 0; the similarity matrix W={w _ij } is a symmetric matrix whose diagonal is 0, and w _ij ∈ [0,1]; f _i , f _j represent the example feature vectors corresponding to i and j example packages respectively, that is, the brightness gradient feature and the color gradient feature vector of the image are synthesized into a 3-dimensional combination vector Mixvector _i ={BrightnessGradient _i , ColorGradient _i }, then Sim(f _i , f _j )=‖Mixvector _i -Mixvector _j ‖ ₂ . In the graph cut framework represented by formula (4), D is an N-dimensional diagonal matrix, and the elements on the diagonal U={U ₁ , U ₂ ,...,U _i ,...,U _j ,...U _N } is the segmentation state vector, and each vector component U _i represents the segmentation state of area i; formula (4 ) represents the visual similarity between region i and region j, and the denominator represents the visual similarity within region i;

Step 22, using the agglomerative hierarchical clustering algorithm to solve the segmentation state vector corresponding to the minimum eigenvalue of R(U), that is, to obtain the optimal segmentation result of the image.

Wherein, the above-mentioned agglomerative hierarchical clustering algorithm refers to step 2 and step 3 of the method with patent application number 201210257591.1.

Test verification

In order to verify the effectiveness of the method of the present invention, the database established by Achanta et al. was used, 300 of which were selected as training images, and the remaining 700 were test images for algorithm verification. Some experimental results are listed as shown in Figure 2. Figure 2 shows the segmentation results of the test image based on the spectrum segmentation algorithm based on multi-scale graph decomposition, the segmentation result of the test image by the method with the patent application number 201210257591.1, and the use of the present invention The split result of the method. described as follows:

In Fig. 2, the subgraphs (a-1) to (a-4) are the original images, the subgraphs (b-1) to (b-4) are the segmentation results of the spectral segmentation algorithm based on multi-scale graph decomposition, and the subgraphs ( c‐1) to (c‐4) are the methods of the patent application number 201210257591.1, and the subfigures (d‐1) to (d‐4) are the methods of the present invention. It can be concluded from the experimental results that when the background is relatively complex, the spectral segmentation algorithm based on multi-scale graph decomposition has serious misclassification and incomplete target segmentation, while the method with the patent application number 201210257591.1 and the method of the present invention have better Segmentation results; when the background is relatively simple and the characteristics of the target are quite different, such as the original image (a-1) and (a-2), the three methods can all segment a relatively complete target; but in the boundary between the target and the background When the transition is slow and the difference is very small, such as the original image (a‐3) and (a‐4), the three methods all have varying degrees of incomplete target segmentation, but the method of the present invention and the patent application number 201210257591.1 The segmentation effect of the method is better, and the method of the present invention can segment more finely at the junction of the target with minimal difference and the background, and can obtain a more accurate segmentation result of the salient target. The input image of the method whose patent application number is 201210257591.1 is the original image, and the coarsening object starts from the pixel level. Although the pixel-level image is relatively fine, in consideration of the amount of calculation, the graph cut method directly adopts the brightness and color features in the When the weight function is defined, only the gray level difference is considered, and the method of the present invention can quickly obtain the salient area marks in the image in combination with the multi-instance learning method, and the example feature vector in each example bag contains the underlying visual features reflecting the target information At the beginning of roughening, the overall features of the image are considered to provide a more accurate segmentation basis for subsequent processing. Therefore, when the transition between the target and the background boundary is slow and the difference is very small, it can still be obtained. good segmentation results. For most of the test images, the number of hierarchical iterations of the method of the present invention is less than that of the method with patent application number 201210257591.1, which greatly reduces the amount of computation and time complexity.

Claims (8)

1. A target segmentation method based on multi-instance learning and graph cut optimization, characterized in that, comprising the steps: Step 1: Use the method of multi-instance learning to model the saliency model on the training image, and use the saliency model to predict the packages and examples in the test image, and obtain the saliency detection result of the test image; specifically include: Step 11, preprocessing the training image, and extracting image brightness gradient features and color gradient features; Step 12, introducing multi-instance learning into the image saliency detection to obtain the saliency detection result of the test image; Step 2: Introduce the saliency detection results of the test image into the graph cut framework, optimize the graph cut framework according to the example feature vector and the label of the example package, solve the suboptimal solution of the graph cut optimization, and obtain the accurate segmentation of the target; Described step 2 specifically comprises the following steps: Step 21, the saliency detection result of the image obtained in step 1 is used as the input of the graph cut algorithm, and the weight function shown in formula (3) is constructed according to the saliency mark of the package and the example feature vector; and the formula (4) is obtained The optimized graph cut cost function shown; In formula (3), w _ij represents the similarity of visual features of the regions corresponding to sample package i and package j, Salien(i) and Salien(j) respectively represent the normalized saliency values of region i and region j, δ In order to adjust the sensitive parameter of visual feature difference, the value is 10-20; the similarity weight between region i and itself is 0; the similarity matrix W={w _ij } is a symmetric matrix whose diagonal is 0, and w _ij ∈ [0,1]; f _i , f _j represent the example feature vectors corresponding to i and j example packages respectively, that is, the brightness gradient feature and the color gradient feature vector of the image are synthesized into a 3-dimensional combination vector Mixvector _i ={BrightnessGradient _i , ColorGradient _i }, then Sim(f _i , f _j )＝||Mixvector _i -Mixvector _j || ₂ ; in the graph-cut framework represented by formula (4), D is an N-dimensional diagonal matrix, and the elements on the diagonal U＝{U ₁ ,U ₂ ,...,U _i ,...,U _j ,...U _N } is the segmentation state vector, Each vector component U _i represents the segmentation state of region i; the numerator of formula (4) represents the distance between region i and region j The visual similarity of , the denominator represents the visual similarity in region i; Step 22, solving the segmentation state vector corresponding to the minimum eigenvalue of R(U), that is, obtaining the optimal segmentation result of the image. 2. The target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 1, wherein, in the step 11, the training image is preprocessed, and the brightness gradient feature and the color gradient feature are extracted, specifically comprising : Step 111, performing color space conversion and quantization preprocessing of each component on the training image to obtain normalized brightness component L and color components a, b; Step 112, calculating the brightness gradient of each pixel in the matrix of the brightness component L; Step 113, calculating the color gradient of each pixel in the matrix of color component a and color component b respectively. 3. The target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 2, wherein said step 111 is specifically as follows: First, gamma correction is performed on the training image to achieve nonlinear adjustment of the image color components, and the training image is converted from the RGB color space to the Lab color space; then the brightness component L of the training image in the Lab color space and the two The color components a and b are normalized to obtain the normalized brightness component L and the color components a and b. 4. The target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 2, wherein said step 112 specifically comprises steps A-D: A. Construct the weight matrix Wights<> of 3 scales; B. Construct the index map matrix Slice_map<> of 3 scales; the index map matrix Slice_map<> of each scale Slice_map<> has the same dimension as the weight matrix Wights<> of the corresponding scale; select 8 directions (0 °, 22.5 °, 45 °, 67.5°, 90°, 112.5°, 135°, 157.5°) divide the matrix into 16 areas, and the value of the elements in each area is the same as the number 0~15 of the area; C. Multiply each index map matrix Slice_map<> with the elements in the weight matrix Wights<> of the corresponding scale one by one to obtain the matrix of the corresponding scale, that is, the neighborhood gradient operator; D. Using a neighborhood gradient operator to calculate the brightness gradient of a pixel to be calculated in the matrix of the brightness component L. 5. the target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 4, is characterized in that, described step A is specifically as follows: Construct the weight matrix Wights<> of three scales respectively; the weight matrix Wights<> is a square matrix with the number of rows and columns equal to 2r+1; the elements in the weight matrix Wights<> are either 0 or 1 , the elements equal to 1 are distributed within the range of the disk with the central element (r+1, r+1) as the center of the square matrix and the radius r as the circle, forming the inscribed circle of the square matrix, and the rest of the elements in the square matrix are 0 ; The 3 scales are r=3, r=5 and r=10 respectively. 6. the target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 4, is characterized in that, described step D is specifically as follows: ① For a certain scale, take a pixel to be sought in the matrix of the luminance component L obtained in step 111 as the center, and perform a calculation with each luminance component within the neighborhood of the pixel to be sought through a neighborhood gradient operator of a certain scale. Dot multiplication to obtain the matrix Neibor<> within the neighborhood of the pixel to be sought; select a straight line in the vertical direction (90°) as the dividing line, and divide the disk in the neighborhood gradient operator into a left semicircle and a right semicircle, The left semicircle includes the 0th sector to the 7th sector, and the right semicircle includes the 8th sector to the 15th sector; the elements of the matrix Neibor<> corresponding to each semicircle form a histogram and normalize it, respectively Recorded as Slice_hist ₁ <> and Slice_hist ₂ <>; H ₁ represents the histogram corresponding to the left semicircle area, H ₂ represents the histogram corresponding to the right semicircle area, i is the value of the bin of the histogram, defined as [0 ,24], that is, the brightness range; ②Calculate the difference between the two normalized histograms through the chi-square distance shown in formula (1), that is, obtain the brightness gradient in the vertical direction of a pixel to be obtained at a certain scale; After calculating the luminance gradient in the vertical direction of a certain scale, select the straight lines in other directions as the dividing line to obtain the luminance gradient in all other directions of a certain scale of the pixel to be obtained; then calculate in the same way as step D Obtain the brightness gradient of the pixel to be requested in all directions on other scales; after completing the calculation of the brightness gradient of the pixel to be requested in all directions of all scales, the final brightness gradient of the pixel to be obtained is calculated by formula (2) : f(x,y,r,n_ori; r=3,5,10; n_ori=1,2,...8)->Brightness Gradient(x,y) (2) In the formula, f is a mapping function, (x, y) is any pixel point to be obtained, r represents the selected scale, n_ori represents the selected direction; Brightness Gradient(x, y) is the pixel point (x, y) The final brightness gradient; the corresponding rule of f is to select the maximum brightness gradient value in each direction in 3 scales as the brightness gradient value in this direction, and sum the brightness gradients in 8 directions to obtain the pixel point (x, y) The final brightness gradient of . 7. The target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 2, wherein in the step 113, the calculation of the color gradient is similar to the calculation of the brightness gradient, and the difference is that the color gradient feature is For the color gradients a and b of the two color components; the selected three scales are r=5, r=10 and r=20; therefore, the sizes of the corresponding weight matrix and map index matrix are 11*11, 21*21 and 41*41; the calculation method of the color gradient of the two color components is the same as that of the brightness gradient, and the final color gradient of each pixel to be obtained in the color component a and b matrix is obtained. 8. The target segmentation method based on multi-instance learning and graph cut optimization as claimed in claim 1, wherein in said step 12, multi-instance learning is introduced into image saliency detection to obtain the saliency detection result of the test image, Specifically include step 121 and step 122: Step 121, using the brightness and color gradient features obtained by the method described in step 11, combined with the multi-instance learning EMDD algorithm to realize the learning of the training set, and obtain a learned saliency detection model; Step 122, substituting the test image into the learned saliency detection model to obtain a saliency detection result of the test image.