CN114821293A - Prototype spectral set generation method based on super-pixels - Google Patents

Abstract

The invention discloses a prototype spectrum set generation method based on superpixels, which belongs to the field of hyperspectral image processing and comprises the following specific steps: firstly, collecting spectral bands for a certain area, and taking each band as an image to form an image set X; then, randomly selecting 40% of wave bands, and averaging pixel points in the image corresponding to each wave band to form an image Y; secondly, calculating the SSIM structural indexes of the images in the image set X and the images Y one by one to serve as scores of the images, and arranging the scores in a descending order; selecting the previous 40% wave band to regenerate the image Y again, and repeatedly calculating the score of each image in the image set X until the wave band is stable; and finally, selecting the first three wave bands with the highest score, inputting the wave bands into a superpixel segmentation algorithm to extract superpixel small blocks at fixed intervals, taking the average spectrum of each superpixel small block as an initial clustering center of a k-means clustering algorithm, and clustering to obtain a prototype spectrum set of the region. The invention reduces the calculation amount in the clustering process and stabilizes the clustering result.

Description

A Prototype Spectral Set Generation Method Based on Superpixels

technical field

The invention belongs to the field of hyperspectral image processing, in particular to a method for generating prototype spectrum sets based on superpixels.

Background technique

In recent years, with the wide application of hyperspectral sensors, hyperspectral image processing technology has also developed, and has played an increasingly important role in military, agricultural and forestry, environmental protection and industrial fields.

In hyperspectral remote sensing technology, the standard spectrum of ground objects is an important basis for hyperspectral remote sensing technology, and the spectral curve is the main basis for band selection and target classification. However, due to various factors such as hyperspectral being greatly affected by light conditions, less labeled hyperspectral images, and slight differences in the standard spectrum of ground objects in different environments, it is difficult to extract the standard spectrum of ground objects effectively.

Superpixels refer to areas composed of adjacent pixels with similar characteristics such as color and brightness. These areas generally do not destroy the boundary information of objects in the image. In hyperspectral images, superpixel patches mainly refer to the homogenous regions in which spectral information is relatively close in hyperspectral images. Introducing the superpixel segmentation algorithm into the components of the prototype spectrum set, on the one hand, it can effectively divide the entire image into nearly pure superpixel blocks containing only similar pixels according to the homogeneity of the pixels; on the other hand, Subsequent calculations are performed by replacing the original pixels with small superpixel blocks, which can greatly reduce the computational complexity of the subsequent clustering algorithm.

Shijin Li and Jianbian Qiu proposed a method for generating prototype spectra in 2013, by clustering hyperspectral images to generate a corresponding prototype spectrum set; on the one hand, this method requires a large amount of calculation, and on the other hand, the results of clustering It is unstable, and it is easy to fall into a local optimum, and it is impossible to obtain an effective prototype spectrum set.

The common area segmentation algorithm that can build superpixels is mainly the k-means clustering method, which directly performs k-means clustering on the hyperspectral image, and uses the stable cluster center obtained by the clustering as the prototype spectral set of the hyperspectral image; However, doing so will lead to high computational time complexity; the cluster centers generated by k-means clustering are unstable, and the results of each experiment are inconsistent, resulting in poor robustness; there is no clear criterion for the number of cluster centers. Multiple experiments are required to get the best results.

SUMMARY OF THE INVENTION

In order to solve the problem of the lack of standard spectrum of ground objects or the difficulty of extracting standard spectra of ground objects, the present invention proposes a method for generating prototype spectrum sets based on superpixels. It keeps the clustering results stable and avoids falling into the local optimum.

The described superpixel-based prototype spectral set generation method, the specific steps are as follows:

Step 1. Use a hyperspectral instrument to collect bands in a certain area, and use each band as an image separately to form a band image set X;

X={x ₁ , x ₂ ,...,x _n ,...,x _N }; N is the total number of bands;

Step 2: Randomly select 40% of the bands, sum the pixels in the image corresponding to each band, and divide by the number of bands, and the average number of pixels obtained forms the image Y;

Step 3, select the images in the image set X one by one, and calculate the SSIM structure index with the image Y respectively, as the corresponding score of each image;

For the image x _n , the formula for calculating the SSIM structure index is as follows:

where μ _X is the mean of the pixels in the current image x _n , μ _Y is the mean of the pixels in the image Y, σ _X is the standard deviation of the pixels in the image x _n , σ _Y is the standard deviation of the pixels in the image Y, C ₁ , C ₂ , and C ₃ represent constants that maintain stability.

Step 4: Arrange in descending order of image scores, select the top 40% bands corresponding to the scores, regenerate image Y using the mean value of all pixel points, and return to step 3 to repeatedly calculate the scores of each image in the image set X until the currently selected band is the same as Same band as last time.

Step 5. Select the first three bands with the highest SSIM structure index score from the current stable bands, input the superpixel segmentation algorithm, and extract the fixed-spaced superpixel small blocks from the segmentation result;

The calculation of the fixed interval is: the total number of superpixel patches divided by the actual number of objects in the area where the image was collected;

Step 6: Use the average spectrum of each superpixel patch at a fixed interval as the initial clustering center of the k-means clustering algorithm, and cluster the superpixel patches divided by all bands in the region to obtain the prototype spectrum set of the region. .

The advantages of the present invention are:

1) A method for generating prototype spectrum sets based on superpixels, in the case of unsupervised, using hyperspectral prototype spectrum sets to replace the standard spectrum of ground objects, which solves the problem that the hyperspectral spectrum is greatly affected by lighting conditions and the ground objects in different environments. The problem of mismatch between the standard spectrum of the ground object and the actual spectral information caused by various factors such as the slight difference in the standard spectrum of the object.

2) A method for generating prototype spectrum sets based on superpixels, which replaces the original spectra of tens of thousands of individual pixels with the average spectrum of small superpixels, which reduces the computational complexity of the clustering process.

3), a prototype spectrum set generation method based on superpixels, the traditional kmeans algorithm is very unstable, and it is easy to fall into the local optimal solution; the present invention evenly distributes the clustering centers on the entire image, and the clustering results are stable. , and basically tend to the optimal solution.

Description of drawings

1 is a flowchart of a method for generating a superpixel-based prototype spectrum set of the present invention;

2 is a schematic diagram of the comparison between the prototype spectral curve calculated by the present invention under two different ground objects and the existing hyperspectral data set Indian_pines;

FIG. 3 is a schematic diagram of the result of superpixel small block segmentation according to the present invention.

Detailed ways

The following is a further detailed explanation of the present invention.

The present invention proposes a flow chart of a method for generating a prototype spectrum set based on superpixels. Since the input of the superpixel segmentation algorithm needs to be consistent with the format of the visible image, that is, only images of three bands are required, while the hyperspectral image has the upper There are hundreds of bands, so the present invention proposes to use the SSIM structure index to extract three bands with the best structure to improve the effect of the superpixel segmentation algorithm; for the problem of unstable cluster centers generated by k-means clustering, the present invention Combined with the label format of SLIC superpixel blocks, it is proposed to use the average spectrum of superpixel blocks with a fixed interval as the measure of the cluster center, so that the clustering results remain stable and avoid falling into the local optimum situation.

The described superpixel-based prototype spectral set generation method is shown in Figure 1, and the specific steps are as follows:

Step 1. Use a hyperspectral instrument to collect bands in a certain area, and use each band as an image separately to form a band image set X;

X={x ₁ , x ₂ ,...,x _n ,...,x _N }; N is the total number of bands;

Step 2: Randomly select 40% of the bands, sum the pixels in the image corresponding to each band, and divide by the number of bands, and the average number of pixels obtained forms the image Y;

Step 3, select the images in the image set X one by one, and calculate the SSIM structure index with the image Y respectively, as the corresponding score of each image;

Due to the special imaging principle of hyperspectral images, that is, different bands are actually images generated from the same region in different narrow bands, the quality of the bands can be determined by the SSIM structure index.

For the image x _n , the formula for calculating the SSIM structure index is as follows:

where μ _X is the mean of the pixels in the current image x _n , μ _Y is the mean of the pixels in the image Y, σ _X is the standard deviation of the pixels in the image x _n , σ _Y is the standard deviation of the pixels in the image Y, C ₁ , C ₂ , and C ₃ represent constants to avoid the denominator being 0 and maintain the stability of the formula. Usually, C ₁ =(K ₁ ×L) ² , C ₂ =(K ₂ ×L) ² , and C ₃ =C ₂ /2 are constants to maintain stability.

K ₁ =0.01, K ₂ =0.03; L _is the pixel range of the image xn.

Step 4: Arrange in descending order of image scores, select the top 40% bands corresponding to the scores, regenerate image Y using the mean value of all pixel points, and return to step 3 to repeatedly calculate the scores of each image in the image set X until the currently selected band is the same as Same band as last time.

This step is mainly to propose the noise band and reduce the influence of the noise band on the screening results.

Step 5. Select the first three bands with the highest SSIM structure index score from the current stable bands, input the superpixel segmentation SLIC algorithm, and after obtaining the corresponding superpixel segmentation result, extract the fixed-spaced superpixel small blocks from the segmentation result;

The calculation of the fixed interval is: the total number of superpixel patches divided by the actual number of objects in the area where the image was collected;

Step 6: Use the average spectrum of each superpixel patch at a fixed interval as the initial clustering center of the k-means clustering algorithm, and cluster the superpixel patches divided by all bands in the region to obtain the prototype spectrum set of the region. .

Aiming at the problem that the cluster centers generated by k-means clustering are unstable; due to the general law of the distribution of ground objects, and the small superpixel blocks generated by the SLIC algorithm are relatively regular, and the order of labels is fixed. Therefore, the present application adopts the average spectrum of small superpixels at fixed intervals as the initial clustering center of the k-means algorithm, and by fixing the number of clustering centers, the problem of unstable clustering results can be effectively solved.

The prototype spectral curve in the prototype spectral set calculated in this application is compared with the public hyperspectral data set Indian_pines, as shown in Figure 2, Figure a is the average spectrum of the ground object with the ground object category of 7 in the Indian_pines data set and its The closest prototype spectral curve, the right picture b is the average spectrum of the ground object with the object category 13 in the Indian_pines dataset and the prototype spectral curve that is closest to it. The solid line is provided in the data according to the real object label. The spectrum curve of the real object, the dotted line is the prototype spectrum curve of the present application; where Figure a shows the largest gap, and Figure b shows the smallest gap. It can be clearly seen that the present application can obtain a better prototype spectrum set that is very close to the spectrum of the real object without using the real object label. (The real object label is difficult to label, or it is time-consuming and laborious to label. The actual object label will not be automatically generated during the actual collection process.)

The result of the superpixel small block segmentation is shown in Figure 3 (the rendering after removing the part of the unmarked features). It can be seen that the application can effectively obtain the arrangement rules, which are compact and can be well described. The superpixel patch of the feature boundary.

Claims (3)

1. A prototype spectral set generation method based on superpixels is characterized by comprising the following specific steps:

firstly, collecting wave bands of a certain area by using a hyperspectral instrument, and forming a wave band image set X by independently using each wave band as an image;

X＝{x ₁ ,x ₂ ,...,x _n ,...,x _N }; n is the total number of wave bands;

then, randomly selecting 40% of wave bands, summing pixel points in the image corresponding to each wave band, dividing the sum by the number of the wave bands to obtain an average pixel number to form an image Y;

then, selecting the images in the image set X one by one, and calculating SSIM structural indexes with the images Y respectively to serve as corresponding scores of all the images; arranging according to the image scores in a descending order, selecting the first 40% wave bands corresponding to the scores, regenerating the image Y by using the average values of all the pixel points, and repeatedly calculating the scores of all the images in the image set X until the currently selected wave band is the same as the last wave band;

finally, selecting the first three wave bands with the highest SSIM structure index score from the current stable wave bands, inputting the wave bands into a superpixel segmentation algorithm, and extracting superpixel small blocks with fixed intervals from the segmentation result; and (3) taking the respective average spectrum of the super-pixel small blocks at fixed intervals as an initial clustering center of a k-means clustering algorithm, and clustering the super-pixel small blocks segmented by all wave bands in the region to obtain a prototype spectrum set of the region.

2. A method of generating a prototype spectral set based on superpixels, according to claim 1, wherein: the image score is calculated by the following process:

for image x _n The formula for calculating the SSIM structural index is as follows:

wherein mu _X Is the current image x _n Mean value of middle pixel, mu _Y Is an imageMean, σ, of pixels in Y _X Is an image x _n Standard deviation, sigma, of the middle pixel _Y Is the standard deviation, C, of the pixels in the image Y ₁ ,C ₂ ,C ₃ Represents a constant to maintain stability.

3. A method of generating a prototype spectral set based on superpixels, according to claim 1, wherein: the fixed interval is calculated as: the total number of super-pixel patches is divided by the actual number of surface features in the area where the image was acquired.

Images