CN107274416A - High spectrum image conspicuousness object detection method based on spectrum gradient and hierarchical structure - Google Patents

Abstract

The invention discloses a hyperspectral image salient target detection method based on spectral gradient and hierarchical structure, which is used to solve the technical problem of large calculation amount in the existing hyperspectral image salient target detection method. The technical solution is to first generate a spectral gradient map; then generate image segmentation regions; establish a saliency detection model based on image hierarchy; then establish a saliency calculation method based on background prior and edge features; and calculate saliency map results. Since the spectral gradient is calculated on the spectral dimension of the original hyperspectral image, the spectral gradient feature of the image is extracted to reduce the adverse effects of uneven illumination. Using the simple linear iterative clustering algorithm (Simple Linear Iterative Clustering, SLIC) to generate superpixels, segment the hyperspectral image and accelerate the calculation process, and measure the significance by calculating the spectral feature contrast between the segmented regions, with a small amount of calculation.

Description

Hyperspectral Image Salient Target Detection Method Based on Spectral Gradient and Hierarchical Structure

technical field

The invention relates to a hyperspectral image salient target detection method, in particular to a hyperspectral image salient target detection method based on spectral gradient and hierarchical structure.

Background technique

Hyperspectral image is the image data obtained by recording the spectral information of various ground objects observed in the field of view by using an imaging spectrometer. With the maturity of hyperspectral imaging technology, imaging equipment has greatly improved its spectral resolution and spatial resolution. As a result, subjects such as object detection, recognition, and tracking, which were originally carried out on conventional images, can gradually be extended to hyperspectral data. Related research on the problem of salient object detection in hyperspectral images is still in the development stage.

The existing hyperspectral image salient target detection methods mainly use the Itti model, which replaces the color features with the spectral features of hyperspectral images, making the model suitable for hyperspectral images. The document "S.L.Moan, A.Mansouri, et al., Saliency for Spectral Image Analysis[J].IEEE Journal of Selected Topics inApplied Earth Observations and Remote Sensing, 2013.6(6):p.2472-2479" discloses a high Spectral image saliency target detection method, which projects the spectrum into the CIELAB color space and utilizes the spectral information by means of image principal component analysis (Principle Component Analysis, PCA). The existing method uses the pixel as the basic unit of significance estimation, and evaluates the difference between the spectra of different pixels by means of principal component analysis, Euclidean distance, and spectral angle (SpectralAngle), so as to measure the difference between each pixel. significant. This method of reflecting the saliency of the whole image by the saliency of the pixel is difficult to avoid the inhomogeneity of the saliency map in the detection result that the object edge response is large and the internal response is very low. In addition, the existing methods all rely on a single model, which cannot eliminate the main difficulties faced in the study of saliency detection in hyperspectral images. There are methods that all depend on a single model. Therefore, it is urgent to break through the inherent ideas in the existing hyperspectral detection methods and propose a new hyperspectral image salient target detection method.

Contents of the invention

In order to overcome the shortage of large amount of calculation in existing methods for detecting salient objects in hyperspectral images, the present invention provides a method for detecting salient objects in hyperspectral images based on spectral gradient and hierarchical structure. The method first generates spectral gradient map; then generates image segmentation region; then establishes a saliency detection model based on image hierarchy; then establishes a saliency calculation method based on background prior and edge features; finally calculates the saliency map result. Since the spectral gradient is calculated on the spectral dimension of the original hyperspectral image, the spectral gradient feature of the image is extracted to reduce the adverse effects of uneven illumination. At the same time, the simple linear iterative clustering algorithm (Simple Linear Iterative Clustering, SLIC) is used to generate superpixels, the hyperspectral image is segmented and the calculation process is accelerated, and the significance is measured by calculating the spectral feature contrast between the segmented regions, with a small amount of calculation .

The technical solution adopted by the present invention to solve the technical problem is: a hyperspectral image salient target detection method based on spectral gradient and hierarchical structure, which is characterized in that it includes the following steps:

Step 1, generating a spectral gradient image.

The spectral gradient is calculated for each pixel to generate a spectral gradient image, so that the spatial relationship in the original image is maintained between the extracted spectral gradient feature vectors.

In the formula, is the jth component of the spectral gradient vector. is the jth component of the original spectral vector. Δλ is the wavelength difference between adjacent bands.

Use the formula (1) to obtain a new spectral gradient data block X for the spectral vector corresponding to each pixel in a hyperspectral data block D.

Step 2, generating image segmentation regions.

Perform a simple linear iterative clustering algorithm on the spectral gradient data block X, the specific steps are as follows:

Input: spectral gradient image X, expected superpixel side length s, weight coefficient m;

Output: mark the segmented image of each superpixel;

1. Initialization process:

1) Initialize a set of initial cluster centers C on the gradient image X with s as the interval length;

2) Adjust each center to the position where the gradient minimum in the 3×3 neighborhood is located;

3) Set the label l _i =-1 corresponding to each pixel, and the distance d _i =+∞ to its current center;

2. Iteratively update the pixel label and each cluster center:

1) For the current clustering center C _k , in a square neighborhood with a side length of 2s, calculate the distance D( _xi ,C _k ) from each pixel x _i to C _k in the neighborhood according to formula (2);

2) If D( _xi ,C _k )<d _i , set the label l _i corresponding to x _i =k, and update d _i =D( _xi ,C _k );

3) Repeat step 1) and step 2) until the change of each center between two iterations before and after is less than the threshold;

In the formula, d _g ( _xi , C _k ) is the Euclidean distance between _xi and C _k in the spectral gradient part. d _s ( _xi , C _k ) is the Euclidean distance between the spatial positions of _xi and C _k . m is the weight coefficient between two distances.

The dual-kernel function is used to replace the density function in the mean shift algorithm to complete the relevant calculations in the mean shift process, and its specific form is

In the formula, x _g is the spectral gradient vector corresponding to the pixel x. x _s is the space coordinate where the pixel x is located. T _g is the bandwidth of the kernel function corresponding to the spectral feature. T _s is the spatial coordinate corresponding to the bandwidth of the kernel function. δ is the normalization coefficient.

The specific steps of the mean shift algorithm are as follows:

Input: superpixel center vector C={C ₁ ,C ₂ ,...,C _k ,...C _n }, spectral threshold T _g , spatial threshold T _s ;

Output: label vector l _sp for the center of the input superpixel;

1) Take the superpixel vector C _k as the initial center to carry out the mean shift process, and record the obtained candidate center C _j ′;

2) For all the samples appearing on the path of C _j ', add 1 to the number of votes for C _j ';

3) Traversing the current set of candidate centers C′, looking for the preferred center C _i ′ whose spectral gradient distance from C _j ′ is less than T _g /2, and whose spatial distance is less than T _s /2;

4) If Merge the vote counts on C _i ′, C _j ′, add the mean value of C _i ′, C _j ′ to C′, and delete C _i ′; otherwise, go to step 5);

5) For each superpixel center, repeat steps 1) to 4) to obtain the final cluster centers;

6) For each superpixel center, take the cluster center C′ _m that received the most votes as its attribution, and obtain l _sp .

Step 3: Establish a salient target detection model based on image hierarchy.

For the spectral threshold T _g required by the sample spectral feature similarity in step 2, and the spatial threshold T _s for controlling the sample neighborhood range, the values are respectively 0.1, 0.2, 0.3, 0.4 times max{r,c}, and 10 , 20, 25, 30. In this way, after the underlying superpixels are clustered at different granularities, a total of 4 levels of clustering results will be generated, forming a 4-layer image hierarchy. Denote the entire superpixel block as the bottom node as in Represents the number of superpixels; in addition, abstract the j-th segmentation region in the i-th layer as a node As mentioned above, superpixels are examined under the hierarchical structure with the number of layers h The final saliency result above, the corresponding saliency detection model is expressed as

In the formula, is to return all the i-th layers containing The node index of . ω _i is the node The weight of the level. is a node Significant value on .

Step 4. Calculation method of position prior, background prior and edge feature saliency.

1. Position prior and background prior.

The mathematical expression of the location prior is as follows:

In the formula, is the Euclidean distance from pixel x _k in region R _i to the center of the image.

The square ring area formed by all the pixels within 10 pixels around the image is selected as the boundary area R _b of the image.

For nodes in the image hierarchy The following three rules are followed when calculating its background prior size:

1) if then yes imposed penalty function

2) Otherwise, if then the penalty function

3) intersection scale The larger the value, the heavier the penalty should be, that is, The absolute value of is larger;

The above three rules make it clear that the contact penalty is calculated Boundary conditions and influencing factors when the background is prior. In the case of satisfying the rules, different specific calculation methods are obtained by using different forms of penalty functions. The penalty function is defined as

In the formula, ξ is the penalty factor carried by each pixel in the boundary area.

2. Significance of edge features.

The mean value of each band of the hyperspectral image is taken as its corresponding grayscale image I _hsi , and the edge features are obtained by using the Canny detector. The regional saliency is calculated by using the edge features on the spatial dimension of the hyperspectral image, and the steps are as follows:

Input: hyperspectral image average grayscale image I _hsi , hierarchy node and its segmentation result graph I _seg ;

output: Marginal feature saliency on

1) Use the Canny detector to extract the edge for I _hsi , and get the result

2) Filter the I _seg with a 3×3 Gaussian filter with a variance of 1.5 to increase the width of the region boundary;

3) Calculate the gradient magnitude image of the filtered I _seg , and obtain the boundary extraction result by binarization

4) Use the following formula to divide the region The image edges on the border are accumulated to get

In the formula, is the split area the boundaries of yes in the Inner edge feature accumulation operation.

Step 5. Calculate the saliency map.

When calculating the saliency map, it is necessary to determine each node in the hierarchy or the division area on each level Significance calculation method. It consists of four parts: spectral gradient area comparison, edge feature saliency, position prior and background prior. The position prior calculation formula is as follows:

When applying the prior, only the part based on the spectral feature area contrast is enhanced, and the part based on the edge feature is not operated. The background prior has a better effect on suppressing the background because the border width of the selected image is small, so the calculation method based on the two features is applied at the same time. Finally, the significance calculation formula is obtained as

In the formula, is the weight coefficient.

The beneficial effects of the present invention are as follows: the method first generates a spectral gradient map; then generates image segmentation regions; then establishes a saliency detection model based on image hierarchy; then establishes a saliency calculation method based on background prior and edge features; finally calculates Salient map results. Since the spectral gradient is calculated on the spectral dimension of the original hyperspectral image, the spectral gradient feature of the image is extracted to reduce the adverse effects of uneven illumination. At the same time, the simple linear iterative clustering algorithm (Simple Linear Iterative Clustering, SLIC) is used to generate superpixels, the hyperspectral image is segmented and the calculation process is accelerated, and the significance is measured by calculating the spectral feature contrast between the segmented regions, with a small amount of calculation .

The present invention will be described in detail below in combination with specific embodiments.

detailed description

The specific steps of the hyperspectral image salient target detection method based on spectral gradient and hierarchical structure in the present invention are as follows:

The hyperspectral remote sensing image has a cubic structure. The spatial dimension reflects the reflectance of pixels corresponding to different positions on the ground in a certain sunlight band, and the spectral dimension reflects the relationship between incident light and reflected light in different bands of pixels at a certain position. A hyperspectral image can be expressed as a p×n data set X _n ={x ₁ ,x ₂ ,...,x _n }, p is the number of bands, n is the total number of pixels in the image; a certain image in the image The element can be expressed as x _i =(x _1i , x _2i ,...,x _pi ) ^T , where x _pi is the reflectivity on the p-th waveband.

Step 1, generating a spectral gradient map.

Spectral Gradient refers to the ratio of the difference between every two adjacent components along the original spectrum vector to the difference of the corresponding wavelength. The vector composed of a series of spectral gradients is called a spectral gradient vector. We refer to the result obtained by calculating the spectral gradient for each pixel as a spectral gradient image, so that the spatial relationship in the original image is maintained between the extracted spectral gradient feature vectors.

In the formula, is the jth component of the spectral gradient vector. is the jth component of the original spectral vector. Δλ is the wavelength difference between adjacent bands.

A new spectral gradient data block X is obtained by using the above formula for the spectral vector corresponding to each pixel in a hyperspectral data block D. The spectral gradient can reduce the brightness difference caused by uneven illumination to a certain extent, and thus can also weaken the influence caused by this difference on the subsequent algorithm steps.

Step 2, generating image segmentation regions.

Since the saliency of each pixel depends on the uniqueness of its features in the neighborhood space (usually in the range of 3x3 pixels), it is difficult for this pixel-level saliency to effectively reflect the saliency of the corresponding macroscopic object. The generation of superpixels effectively reduces redundant information in local areas of images, simplifies image expression, and reduces the complexity of subsequent image processing tasks; in addition, superpixels are conducive to extracting visual information in the middle layer of images, which is more in line with human perception of images. A simple linear iterative clustering algorithm (Simple Linear Iterative Clustering, SLIC) is performed on the spectral gradient data X. The specific steps of the SLIC algorithm are as follows:

Input: spectral gradient image X, expected superpixel side length s, weight coefficient m;

Output: mark the segmented image of each superpixel;

1. Initialization process:

1) Initialize a set of initial cluster centers C on the gradient image X with s as the interval length;

2) Adjust each center to the position where the gradient minimum in the 3×3 neighborhood is located;

3) Set the label l _i =-1 corresponding to each pixel, and the distance d _i =+∞ to its current center;

2. Iteratively update the pixel label and each cluster center:

1) For the current clustering center C _k , in a square neighborhood with a side length of 2s, calculate the distance D( _xi ,C _k ) from each pixel x _i to C _k in the neighborhood according to formula (2);

2) If D( _xi ,C _k )<d _i , set the label l _i corresponding to x _i =k, and update d _i =D( _xi ,C _k );

3) Repeat step 1) and step 2) until the change of each center between two iterations before and after is less than the threshold;

In the formula, d _g ( _xi , C _k ) is the Euclidean distance between _xi and C _k in the spectral gradient part. d _s ( _xi , C _k ) is the Euclidean distance between the spatial positions of _xi and C _k . m is the weight coefficient between two distances.

The above SLIC superpixel generation algorithm can divide the spectral gradient image into many partitions with relatively small spatial dimensions. The result of superpixel segmentation is an over-segmentation result. Objects in real scenes are often segmented into very small areas, which is not conducive to the salient target detection algorithm to obtain more uniform results inside the object. Therefore, superpixels are clustered to obtain higher-level visual features and improve the internal uniformity of detection algorithm results. However, since the sample to be clustered in the present invention is the center of each superpixel, which contains spectral features and spatial coordinates, the present invention uses a dual-kernel function to replace the density function in the mean-shift algorithm (Mean-Shift) to complete the mean shift The relevant calculation in the process, its specific form is

In the formula, x _g is the spectral gradient vector corresponding to the pixel x. x _s is the space coordinate where the pixel x is located. T _g is the Kernel Bandwidth corresponding to the spectral feature. T _s is the spatial coordinate corresponding to the bandwidth of the kernel function. δ is the normalization coefficient.

The specific steps of the Mean-Shift algorithm are as follows:

Input: superpixel center vector C={C ₁ ,C ₂ ,...,C _k ,...C _n }, spectral threshold T _g , spatial threshold T _s ;

Output: label vector l _sp for the center of the input superpixel;

1) Take the superpixel vector C _k as the initial center to carry out the mean shift process, and record the obtained candidate center C _j ′;

2) For all the samples appearing on the path of C _j ', add 1 to the number of votes for C _j ';

3) Traversing the current set of candidate centers C′, looking for the preferred center C _i ′ whose spectral gradient distance from C _j ′ is less than T _g /2, and whose spatial distance is less than T _s /2;

4) If Merge the vote counts on C _i ′, C _j ′, add the mean value of C _i ′, C _j ′ to C′, and delete C _i ′; otherwise, go to step 5);

5) For each superpixel center, repeat steps 1) to 4) to obtain the final cluster centers;

6) For each superpixel center, take the cluster center C′ _m that has received the most votes as its attribution, and get l _sp ;

Step 3: Establish a salient target detection model based on image hierarchy.

After completing the above steps, if only the spatial relationship between the segmented regions is considered on a single segmentation level, the relationship between them and the Semantic Region cannot be well described, so that such as saliency detection and image Processing tasks related to semantic understanding are difficult to achieve good results. In fact, multiple segmentation levels of an image jointly constitute its hierarchical structure, and its effective use can improve the utilization of image semantic information by enhanced processing methods.

For the spectral threshold T _g required by the sample spectral feature similarity in the previous step, and the spatial threshold T _s for controlling the sample neighborhood range, the values are respectively 0.1, 0.2, 0.3, and 0.4 times max{r,c}, and 10, 20, 25, 30. In this way, after the underlying superpixels are clustered at different granularities, a total of 4 levels of clustering results will be generated, forming a 4-layer image hierarchy. Denote the entire superpixel block as the bottom node as in Represents the number of superpixels; in addition, abstract the j-th segmentation region in the i-th layer as a node As mentioned above, superpixels are examined under the hierarchical structure with the number of layers h On the final saliency result, the corresponding saliency detection model can be expressed as

In the formula, is to return all the i-th layers containing The node index of . ω _i is the node The weight of the level. is a node Significant value on .

Step 4. Calculation method of position prior, background prior and edge feature saliency.

Considering that human eyes tend to pay more attention to the center of the scene than the surrounding areas, the present invention introduces a priori based on the position of the segmented area and background priori. The background prior also assumes that salient regions tend to be closer to the center of the image. So regions distributed near image boundaries are usually more likely to be background parts, and penalties are imposed on these boundary regions to suppress their response size to detection methods. The invention uses the SLIC algorithm and the Mean-Shift algorithm to construct an image hierarchical structure on a spectral gradient image, and provides a salient target detection model under the hierarchical structure. And a new saliency calculation method based on edge features and background prior is introduced to further improve the model's ability to describe regional saliency. Next, the position prior, background prior and edge feature saliency calculation methods will be introduced in detail.

1. Position prior and background prior.

Considering that the human eye tends to pay more attention to the center of the scene than the surrounding areas, the mathematical expression of the position prior is as follows:

In the formula, is the Euclidean distance from pixel x _k in region R _i to the center of the image.

In the present invention, the square ring area formed by all pixels within 10 pixels from the surrounding image is selected as the boundary area R _b of the image. To suppress segmented regions that intersect with _Rb , we employ a “contact penalty” approach to achieve this. For nodes in the image hierarchy When calculating its background prior size, the following three rules should be followed:

1) if then yes imposed penalty function

2) Otherwise, if then the penalty function

3) intersection scale The larger the value, the heavier the penalty should be, that is, The absolute value of is larger;

The above three rules make it clear that the contact penalty is calculated Boundary conditions and influencing factors when the background is prior. In the case of satisfying the rules, different specific calculation methods can be obtained by using different forms of penalty functions. The penalty function is defined as

In the formula, ξ is the penalty factor carried by each pixel in the boundary area.

2. Significance of edge features.

The human visual system is more sensitive to the edge of the image, and the visual attention is easily attracted to the image area with obvious edge features, mainly because the edge of the image is usually the place where the gray level of the pixel changes sharply. Therefore, the present invention not only uses the region comparison method based on the spectral gradient feature, but also introduces the edge feature in the spatial dimension of the hyperspectral image to further improve the effect. The present invention takes the average value of each band of the hyperspectral image as its corresponding grayscale image I _hsi , and uses Canny detectors to obtain edge features. The present invention utilizes the edge features on the space dimension of the hyperspectral image to calculate the regional saliency, and the steps are as follows:

Input: hyperspectral image average grayscale image I _hsi , hierarchy node and its segmentation result graph I _seg ;

output: Marginal feature saliency on

1) Use the Canny detector to extract the edge for I _hsi , and get the result

2) Filter the I _seg with a 3×3 Gaussian filter with a variance of 1.5 to increase the width of the region boundary;

3) Calculate the gradient magnitude image of the filtered I _seg , and obtain the boundary extraction result by binarization

4) Use the following formula to divide the region The image edges on the border are accumulated to get

In the formula, is the split area the boundaries of yes in the Inner edge feature accumulation operation.

Step 5. Calculate the saliency map.

When calculating the saliency map, it is mainly to determine each node in the hierarchical structure or the division area on each level Significance calculation method. As mentioned earlier, in this chapter method It mainly consists of four parts: spectral gradient area comparison, edge feature saliency, position prior and background prior. The position prior calculation formula is as follows:

When applying the prior, considering that the weighting of the distance by the position prior with an exponential function has a strong enhancement effect on the central area, it is not completely beneficial to improve the performance of the detection method; therefore, only the part based on the comparison of spectral feature areas is enhanced. No operation is performed on the part based on edge features. The background prior has a better effect on suppressing the background because the border width of the selected image is small, so the calculation method based on the two features is applied at the same time. Finally, the significance calculation formula of the present invention is obtained as

In the formula, is the weight coefficient.

Claims (1)

1. a hyperspectral image salient target detection method based on spectral gradient and hierarchical structure, is characterized in that comprising the following steps: Step 1, generating a spectral gradient image; Calculate the spectral gradient for each pixel to generate a spectral gradient image, so that the spatial relationship in the original image is maintained between the extracted spectral gradient feature vectors; <mrow><msubsup><mi>g</mi><mi>i</mi><mrow><mo>(</mo><mi>j</mi><mo>)</mo></mrow></msubsup><mo>=</mo><mfrac><mn>1</mn><mrow><mi>&amp;Delta;</mi><mi>&amp;lambda;</mi></mrow></mfrac><mo>{</mo><msubsup><mi>y</mi><mi>i</mi><mrow><mo>(</mo><mi>j</mi><mo>)</mo></mrow></msubsup><mo>-</mo><msubsup><mi>y</mi><mi>i</mi><mrow><mo>(</mo><mi>j</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup><mo>}</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow></mrow> In the formula, is the jth component of the spectral gradient vector; is the jth component of the original spectral vector; Δλ is the wavelength difference of adjacent bands; Use the formula (1) to obtain a new spectral gradient data block X for the spectral vector corresponding to each pixel in a hyperspectral data block D; Step 2, generating image segmentation regions; Perform a simple linear iterative clustering algorithm on the spectral gradient data block X, the specific steps are as follows: Input: spectral gradient image X, expected superpixel side length s, weight coefficient m; Output: mark the segmented image of each superpixel; 1. Initialization process: 1) Initialize a set of initial cluster centers C on the gradient image X with s as the interval length; 2) Adjust each center to the position where the gradient minimum in the 3×3 neighborhood is located; 3) Set the label l _i =-1 corresponding to each pixel, and the distance d _i =+∞ to its current center; 2. Iteratively update the pixel label and each cluster center: 1) For the current clustering center C _k , in a square neighborhood with a side length of 2s, calculate the distance D( _xi ,C _k ) from each pixel x _i to C _k in the neighborhood according to formula (2); 2) If D( _xi ,C _k )<d _i , set the label l _i corresponding to x _i =k, and update d _i =D( _xi ,C _k ); 3) Repeat step 1) and step 2) until the change of each center between two iterations before and after is less than the threshold; <mrow><mi>D</mi><mrow><mo>(</mo><mrow><msub><mi>x</mi><mi>i</mi></msub><mo>,</mo><msub><mi>C</mi><mi>k</mi></msub></mrow><mo>)</mo></mrow><mo>=</mo><msqrt><mrow><msubsup><mi>d</mi><mi>g</mi><mn>2</mn></msubsup><mrow><mo>(</mo><mrow><msub><mi>x</mi><mi>i</mi></msub><mo>,</mo><msub><mi>C</mi><mi>k</mi></msub></mrow><mo>)</mo></mrow><mo>+</mo><msup><mrow><mo>&amp;lsqb;</mo><mfrac><mrow><msub><mi>d</mi><mi>s</mi></msub><mrow><mo>(</mo><mrow><msub><mi>x</mi><mi>i</mi></msub><mo>,</mo><msub><mi>C</mi><mi>k</mi></msub></mrow><mo>)</mo></mrow></mrow><mi>s</mi></mfrac><mo>&amp;rsqb;</mo></mrow><mn>2</mn></msup><msup><mi>m</mi><mn>2</mn></msup></mrow></msqrt><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow></mrow> In the formula, d _g ( _xi , C _k ) is the Euclidean distance between _xi and C _k in the spectral gradient; d _s ( _xi , C _k ) is the Euclidean distance between _xi and C _k ; m is The weight coefficient between two distances; The dual-kernel function is used to replace the density function in the mean shift algorithm to complete the relevant calculations in the mean shift process, and its specific form is <mrow><msub><mi>K</mi><mrow><msub><mi>T</mi><mi>s</mi></msub><mo>,</mo><msub><mi>T</mi><mi>g</mi></msub></mrow></msub><mrow><mo>(</mo><mi>x</mi><mo>)</mo></mrow><mo>=</mo><mfrac><mi>&amp;delta;</mi><mrow><msubsup><mi>T</mi><mi>s</mi><mn>2</mn></msubsup><msubsup><mi>T</mi><mi>g</mi><mn>2</mn></msubsup></mrow></mfrac><mi>k</mi><mrow><mo>(</mo><mo>|</mo><mo>|</mo><mfrac><msub><mi>x</mi><mi>s</mi></msub><msub><mi>T</mi><mi>s</mi></msub></mfrac><mo>|</mo><msup><mo>|</mo><mn>2</mn></msup><mo>)</mo></mrow><mi>k</mi><mrow><mo>(</mo><mo>|</mo><mo>|</mo><mfrac><msub><mi>x</mi><mi>g</mi></msub><msub><mi>T</mi><mi>g</mi></msub></mfrac><mo>|</mo><msup><mo>|</mo><mn>2</mn></msup><mo>)</mo></mrow><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>3</mn><mo>)</mo></mrow></mrow> In the formula, x _g is the spectral gradient vector corresponding to the pixel x; x _s is the spatial coordinate of the pixel x; T _g is the bandwidth of the kernel function corresponding to the spectral feature; T _s is the bandwidth of the kernel function corresponding to the spatial coordinate; conversion coefficient; The specific steps of the mean shift algorithm are as follows: Input: superpixel center vector C={C ₁ ,C ₂ ,...,C _k ,...C _n }, spectral threshold T _g , spatial threshold T _s ; Output: label vector l _sp for the center of the input superpixel; 1) Take the superpixel vector C _k as the initial center to carry out the mean shift process, and record the obtained candidate center C′ _j ; 2) For all the samples appearing on the formation path of C′ _j , add 1 to the number of votes for C′ _j ; 3) Traversing the current set of candidate centers C′, looking for the preferred center C′ _i whose spectral gradient distance from C′ _j is smaller than T _g /2, and whose spatial distance is smaller than T _s /2; 4) if Then merge the vote counts on C′ _i and C′ _j , and add the average value of C′ _i and C′ _j to C′ at the same time, and delete C′ _i ; otherwise, go to step 5); 5) For each superpixel center, repeat steps 1) to 4) to obtain the final cluster centers; 6) For each superpixel center, take the cluster center C′ _m that has received the most votes as its attribution, and get l _sp ; Step 3, establishing a salient target detection model based on image hierarchy; For the spectral threshold T _g required by the sample spectral feature similarity in step 2, and the spatial threshold T _s for controlling the sample neighborhood range, the values are respectively 0.1, 0.2, 0.3, 0.4 times max{r,c}, and 10 , 20, 25, 30; in this way, after the underlying superpixels are clustered at different granularities, a total of 4 levels of clustering results will be generated, forming a 4-layer image hierarchy; the superpixel block will be used as the underlying node All written as in Represents the number of superpixels; in addition, abstract the j-th segmentation region in the i-th layer as a node As mentioned above, superpixels are examined under the hierarchical structure with the number of layers h The final saliency result above, the corresponding saliency detection model is expressed as In the formula, is to return all the i-th layers containing The node subscript of ; ω _i is the node The weight of the level; is a node Significant value on ; Step 4, position prior, background prior and edge feature saliency calculation method; 1. Position prior and background prior; The mathematical expression of the location prior is as follows: <mrow><mi>p</mi><mrow><mo>(</mo><msub><mi>R</mi><mi>i</mi></msub><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mrow><mi>&amp;omega;</mi><mrow><mo>(</mo><msub><mi>R</mi><mi>i</mi></msub><mo>)</mo></mrow></mrow></mfrac><munder><mo>&amp;Sigma;</mo><mrow><msub><mi>x</mi><mi>k</mi></msub><mo>&amp;Element;</mo><msub><mi>R</mi><mi>i</mi></msub></mrow></munder><mi>exp</mi><mrow><mo>(</mo><mo>-</mo><mn>2</mn><msubsup><mi>d</mi><msub><mi>x</mi><mi>k</mi></msub><mn>2</mn></msubsup><mo>)</mo></mrow><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>5</mn><mo>)</mo></mrow></mrow> In the formula, is the Euclidean distance from the pixel x _k in the region R _i to the center of the image; Select the square ring area formed by all pixels within 10 pixels around the image as the boundary area R _b of the image; for the nodes in the image hierarchy The following three rules are followed when calculating its background prior size: 1) if then yes imposed penalty function 2) Otherwise, if then the penalty function 3) Intersection scale with R _b The larger the value, the heavier the penalty should be, that is, The absolute value of is larger; The above three rules make it clear that the contact penalty is calculated Boundary conditions and influencing factors when the background is prior; in the case of satisfying the rules, different forms of penalty functions are used to obtain different specific calculation methods; the penalty function is defined as <mrow><mi>&amp;kappa;</mi><mrow><mo>(</mo><msub><mover><mi>N</mi><mo>^</mo></mover><mrow><mi>i</mi><mo>,</mo><mi>j</mi></mrow></msub><mo>)</mo></mrow><mo>=</mo><mi>&amp;xi;</mi><mo>&amp;CenterDot;</mo><mo>|</mo><msub><mover><mi>N</mi><mo>^</mo></mover><mrow><mi>i</mi><mo>,</mo><mi>j</mi></mrow></msub><mo>&amp;cap;</mo><msub><mi>R</mi><mi>b</mi></msub><mo>|</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>6</mn><mo>)</mo></mrow></mrow> 2 In the formula, ξ is the penalty factor carried by each pixel in the boundary area; 2. Significance of edge features; The mean value of each band of the hyperspectral image is used as its corresponding grayscale image I _hsi , and the edge features are obtained by using the Canny detector; the regional saliency is calculated by using the edge features on the spatial dimension of the hyperspectral image, and the steps are as follows: Input: hyperspectral image average grayscale image I _hsi , hierarchy node and its segmentation result graph I _seg ; output: Marginal feature saliency on 1) Use the Canny detector to extract the edge for I _hsi , and get the result 2) Filter the I _seg with a 3×3 Gaussian filter with a variance of 1.5 to increase the width of the region boundary; 3) Calculate the gradient magnitude image of the filtered I _seg , and obtain the boundary extraction result by binarization 4) Use the following formula to divide the region The image edges on the border are accumulated to get In the formula, is the split area the boundaries of yes in the The edge feature accumulation operation within; Step five, calculate the saliency map; When calculating the saliency map, it is necessary to determine each node in the hierarchy or the division area on each level Significance calculation method; It consists of four parts: spectral gradient area comparison, edge feature salience, position prior, and background prior; the position prior calculation formula is as follows: <mrow><mi>p</mi><mrow><mo>(</mo><msub><mover><mi>N</mi><mo>^</mo></mover><mrow><mi>i</mi><mo>,</mo><mi>j</mi></mrow></msub><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mrow><mo>|</mo><msub><mover><mi>N</mi><mo>^</mo></mover><mrow><mi>i</mi><mo>,</mo><mi>j</mi></mrow></msub><mo>|</mo></mrow></mfrac><munder><mo>&amp;Sigma;</mo><mrow><msub><mi>x</mi><mi>k</mi></msub><mo>&amp;Element;</mo><msub><mover><mi>N</mi><mo>^</mo></mover><mrow><mi>i</mi><mo>,</mo><mi>j</mi></mrow></msub></mrow></munder><mi>exp</mi><mo>{</mo><mo>-</mo><mn>2</mn><msubsup><mi>d</mi><msub><mi>x</mi><mi>k</mi></msub><mn>2</mn></msubsup><mo>}</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>8</mn><mo>)</mo></mrow></mrow> When applying the prior, only the part based on the comparison of the spectral feature area is enhanced, and the part based on the edge feature is not operated; the background prior has a better effect on suppressing the background because the selected image boundary width is small, so for Calculation methods based on the two features are applied simultaneously; the final significance calculation formula is In the formula, is the weight coefficient.