Hyperspectral image unsupervised segmentation method based on mean shift

Abstract

The invention discloses a hyperspectral influence unsupervised segmentation method based on mean shift clustering. According to the invention, unsupervised semantic segmentation is carried out on a hyperspectral remote sensing image by using mean shift iteration and an improved linear clustering algorithm. Existing hyperspectral image segmentation generally has two problems that firstly, a data set required by a supervised classification method is insufficient; and 2, an unsupervised method has a relatively high requirement on priori knowledge. According to the hyperspectral influence unsupervised segmentation method based on mean shift clustering, a mean shift iteration method is referred and combined, a linear clustering method is improved, and the method does not need the number of segmentation classes as an input parameter and does not need priori knowledge about an image and a research area.

Description

Technical Field

[0001] The present invention relates to the field of pixel classification in remote sensing image processing, and in particular to an unsupervised semantic segmentation technology based on hyperspectral images. Background Art

[0002] By analyzing the specific electromagnetic range of hyperspectral remote sensing images, we can determine the material and state of the observed objects based on their spectral characteristics, and distinguish buildings, farmland, rivers and other objects in the image based on their spectral distribution. Therefore, pixel-level classification of hyperspectral remote sensing images is of great significance in the fields of environmental monitoring, land use planning and ecological protection.

[0003] In order to effectively use hyperspectral images for object classification, existing methods usually use data-dependent neural network methods for supervised classification. Such methods require a huge amount of image datasets with precise labels. However, labeling is an interactive task that takes a lot of time and effort. At the same time, the number of publicly available hyperspectral datasets is limited, and they are usually composed of a single image. The data volume is very small and cannot meet the needs of engineering practice. On the other hand, for semantic segmentation problems, although unsupervised segmentation methods can be appropriately introduced, these methods are usually limited to using general label clustering instead of specific image objects, and it is still necessary to specify limited object types for objects in the image.

[0004] This paper designs an unsupervised segmentation method for hyperspectral images based on mean shift, which uses the mean shift algorithm to perform unsupervised semantic segmentation on hyperspectral remote sensing images to complete the classification of pixels in the image. Unlike existing methods, this method does not require the number of segmentation classes as an input parameter, nor does it require prior knowledge about the image and the study area. Summary of the invention

[0005] To overcome the above problems encountered in the classification of hyperspectral image pixels, this paper designs an unsupervised segmentation method for hyperspectral images based on mean shift. This technique does not require the number of segmentation classes as an input parameter, nor does it require prior knowledge about the image and the study area.

[0006] The technical solution adopted by the present invention is:

[0007] Step 1: Image normalization

[0008] Step 2: Pixel similarity information extraction

[0009] Step 3: Calculate the cost distance and perform simple linear clustering

[0010] Step 4: Mean shift clustering iteration

[0011] Step 5: Iteration terminates and clustering results are output

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] (1) This technology does not require the number of segmentation classes as an input parameter during segmentation;

[0014] (2) This technology does not require prior knowledge of the image and the study area during segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Attached Figure 1 :Flowchart of unsupervised segmentation algorithm for hyperspectral images based on mean shift

[0016] Attached Figure 2 :Simple linear clustering process diagram

[0017] Attached Figure 3 :Schematic diagram of mean shift principle

[0018] Attached Figure 4 :Schematic diagram of algorithm effect demonstration DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the implementation modes and the accompanying drawings.

[0020] The algorithm flow of the present invention is as follows Figure 1 As shown, the specific method implemented by the present invention comprises the following steps:

[0021] Step 1: Image normalization processing

[0022] In order to avoid outliers and noise in the hyperspectral signal, we first consider the value range of 95% of the image pixels and perform limited normalization to map the values ​​of all pixels to [0, V]. The normalization process is shown in Formula 1:

[0023]

[0024] In the formula, p i Indicates the pixel value of the current position, p max Indicates the maximum value of the pixel value, p min Indicates the minimum value of the pixel values.

[0025] Step 2: Pixel similarity information extraction

[0026] The similarity information of the normalized pixels is extracted, and the optimal number of clusters suitable for the input data is found through iteration. Each cluster center is defined by the average value of the hyperspectral image pixel value assigned to the u-th cluster, and the relationship is shown in Formula 2:

[0027]

[0028] In the formula, Q u represents the u-th cluster center, It represents the average value of the i-th group of pixels in the u-th class, and b represents the band channel.

[0029] Once the algorithm finds a clustering solution, the spectral information of each pixel in the image is connected to the cluster center to which it belongs, using F i To describe the position (x i ,y i ) can be expressed by formula 3:

[0030] F i = <P i ,Q u ,x i ,y i > (3)

[0031] In the formula, F i Indicates the position (x i ,y i ) at pixel P i and the u-th cluster center Q u The relationship vector of this vector is defined by Formula 4:

[0032] Size(F i )=2×b+2 (4)

[0033] In the formula, Size(F i ) represents the size of the vector, and b represents the number of image band channels.

[0034] Step 3: Calculate the cost distance and perform simple linear clustering

[0035] The present invention adopts an improved simple linear iterative clustering algorithm to form superpixels. The principle diagram of linear clustering in this step is as follows: Figure 2 As shown. Assume there are K superpixel cluster centers C k , determined by a fixed grid spacing S, the value of S is defined by Formula 5:

[0036]

[0037] In the formula, S represents the grid spacing of the superpixel cluster, K represents the number of superpixel centers, and N represents the total number of pixels. After the grid is initialized, an iterative search is performed starting from the cluster center to convert each pixel P i Assign to the closest superpixel, and the search area is the 2S×2S area around the cluster center. The superpixel assignment method is determined by the minimum cost distance, which is calculated as shown in Formula 6:

[0038]

[0039]

[0040]

[0041]

[0042] Where D s represents the cost distance used in the present invention, d spec Represents the band distance component, d clust Represents the pixel distance component, d xy represents the spatial distance component, L is the total number of pixels, k is the number of superpixels, b is the number of bands, p is the pixel value, x and y are the coordinates of the pixel respectively, and i is the current pixel.

[0043] Step 4: Mean shift clustering iteration

[0044] The pixel clustering is iterated by the mean shift method. The mean shift means that in each round of iteration, the cluster center is moved toward the direction of the sum of the cluster vectors until the sum of the cluster vectors meets the conditions. The principle is as follows: Figure 3 As shown, the mean shift vector can be obtained by formula 7:

[0045]

[0046] Where M represents the mean offset vector at each iteration, F i represents the relationship vector in step 2, i represents the i-th vector, and K represents the number of cluster centers.

[0047] Step 5: Terminate the iteration and output the clustering effect

[0048] When the mean offset vector is smaller than the threshold, that is, when the normalized result of the global offset is smaller than m=0.2, the iteration is stopped and the clustering effect is output.

[0049] Figure 4 This is a schematic diagram of the segmentation effect shown in the present invention. In the figure, the pixels corresponding to the typical objects are well segmented, and the recall rate reaches 90%.

[0050] The above description is only a specific implementation mode of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all disclosed features, or all methods or steps in the process, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. An unsupervised segmentation method for hyperspectral images based on mean shift clustering, characterized in that: The following steps are involved: Step 1: Image normalization processing; Step 2: Pixel similarity information extraction; Step 3: This step is the core content of the patent; calculate the cost distance and perform simple linear clustering; Step 4: This step is the core content of the patent; the clustering algorithm is iterated through mean shift. Step 5: Set the iteration termination condition, terminate the iteration, and output the clustering results.

2. The method according to claim 1, characterized in that: Use in step 2 Calculate the cost distance.

3. The method according to claim 1, characterized in that: In step 3, the mean shift algorithm is used to iteratively cluster the results.

Images