Target identification method based on hyperspectral remote sensing and machine learning fusion

By establishing a standard image database and preprocessing of hyperspectral remote sensing images, combining PHash algorithm and hierarchical clustering analysis, the problems of low recognition accuracy and large calculation volume in the existing technology are solved, and the rapid and accurate goals are achieved. Identify.

CN118941782AInactive Publication Date: 2024-11-12TIANZE SMART TECH (CHENGDU) CO LTD
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411420467.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-12
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the prior art, the object recognition method based on hyperspectral remote sensing and machine learning is too narrow in application range and has low recognition accuracy, which leads to low recognition efficiency and reduced recognition speed.

Method used

By establishing a standard image database, multiple feature images containing multiple reference targets, noise reduction, dimensionality reduction and feature extraction of hyperspectral remote sensing images are performed, feature comparison and analysis is performed using PHash algorithm, and the unknown is combined with hierarchical clustering analysis algorithm. The targets are analyzed and entered into the database to achieve rapid identification.

Benefits of technology

It realizes rapid identification of various different goals, improves recognition accuracy and efficiency, reduces the amount of calculation in the feature extraction process, and enhances the flexibility and diversity of the identification system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941782A_ABST
    Figure CN118941782A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of target recognition, and provides a target recognition method based on hyperspectral remote sensing and machine learning fusion, which comprises the following steps: establishing a standard image database; obtaining a plurality of hyperspectral remote sensing images of the target area; preprocessing and classifying each hyperspectral remote sensing image; performing feature extraction on each type of hyperspectral remote sensing images to obtain a plurality of target feature images corresponding to each type of hyperspectral remote sensing images; performing comparative analysis on the shape feature of each target feature image and the shape feature of each feature image based on a PHhash algorithm to obtain an analysis result; processing the analysis result by adopting a hierarchical clustering analysis algorithm, and outputting a target recognition result; the target recognition result comprises a known target and an unknown target. The method has the advantages of wide application range, multiple types of identified targets, and accurate and rapid identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of target recognition, and in particular relates to a target recognition method based on the fusion of hyperspectral remote sensing and machine learning. Background Art

[0002] The imaging spectrometer with high spectral resolution provides a very narrow imaging band for each imaging pixel, and it is continuously distributed in a certain spectral interval. The spectral curve of the ground object obtained by the high-resolution sensor is a continuous spectral signal. Therefore, the ability to distinguish and identify ground objects is greatly improved, and different categories of the same ground object can be distinguished. The phenomenon of "same object, different spectrum" and "same spectrum, different objects" is reduced, and the imaging channels are greatly increased. In the analysis of different applications, the selectivity of the spectrum becomes flexible and diversified, which greatly increases the number of targets analyzed by remote sensing means and realizes the identification of different targets.

[0003] In the prior art, Chinese invention patent CN117671469A discloses a method for automatic recognition of very small seabed targets based on hyperspectral images and machine learning. However, this method involves specific fields and has a narrow application scope. The recognition accuracy is not high when applied to the recognition of other targets.

[0004] Chinese invention patent CN115760814A discloses a remote sensing image fusion method and system based on a dual-coupled deep neural network, which requires inputting hyperspectral remote sensing images and multispectral remote sensing images into a dual-coupled deep neural network model for image fusion to achieve target recognition. This will greatly increase the amount of calculation required, resulting in low recognition efficiency and reduced recognition speed.

[0005] In view of this, this application is hereby filed. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention is implemented by the following technical solutions: In a first aspect of the present invention, a target recognition method based on the fusion of hyperspectral remote sensing and machine learning is provided, comprising the following steps: S00, establishing a standard image database; the standard image database contains multiple feature images of multiple reference targets; each feature image contains shape features and contour features of the reference target; S10, obtaining multiple hyperspectral remote sensing images of the target area; S20, performing noise reduction processing on each hyperspectral remote sensing image, performing dimensionality reduction processing on each hyperspectral remote sensing image after the noise reduction processing; performing image classification on the multiple hyperspectral remote sensing images after the dimensionality reduction processing; S30, performing feature extraction on each type of hyperspectral remote sensing image to obtain a plurality of target feature images corresponding to each type of hyperspectral remote sensing image; each target feature image contains shape features and contour features of the target to be identified; S40, comparing and analyzing the feature image generated in the previous step with the image in the preset standard database based on the PHash algorithm, and dividing them into known targets and unknown targets; S50, analyzing the above unknown target features through a hierarchical clustering analysis algorithm, and entering the features into a database to complete the learning.

[0007] In one embodiment of the present invention, S20 includes: S201, decomposing the image into high-frequency components and low-frequency components through a global sparse regularization term; S202, selecting sample values ​​of smoothing parameters from high-frequency components by block division, then fitting a smoothing parameter surface using a Bessel method, and then calculating corresponding smoothing parameter values ​​according to the coordinates of the pixel points; S203, based on the theory that high-frequency components and low-frequency components are highly correlated, the low-frequency image is smoothed using a smoothing parameter obtained from the high-frequency component, and an image to be classified is obtained by iterative optimization; S204: classify different targets in the above-mentioned image to be classified according to preset standard values.

[0008] In one embodiment of the present invention, S30 includes: S301, constructing a feature descriptor; S302, with the feature point as the center, construct multiple discretized Bresenham concentric circles of different radii, and then obtain N sampling points with the same spacing on each concentric circle. Since this neighborhood sampling mode will produce image grayscale aliasing during sampling, the standard deviation of the Gaussian function used is proportional to the spacing between points on each concentric circle.

[0009] S303, then rotating the sampling pattern around the feature point, so that the feature descriptor has rotation invariance.

[0010] S304: Form a binary character string descriptor of the feature point.

[0011] S305. Select feature points with larger RGB value difference, brightness and contrast from the descriptor obtained in the above manner, introduce directional gradient features and local structural features of the feature points, generate a binary comparison template, and calculate features of the area around the feature points, wherein the feature points with larger RGB value difference, brightness and contrast are feature points whose RGB value difference is higher than a preset threshold of the RGB value difference, whose brightness is higher than a preset threshold of the brightness, and whose contrast is higher than a preset threshold of the contrast.

[0012] In one embodiment of the present invention, S40 includes: S401, converting the picture into a grayscale image; S402, calculating the DCT transform of the image to obtain a 32x32 DCT coefficient matrix; S403, retaining the 8x8 matrix in the upper left corner, and calculating the mean value of DCT; S404, according to the 8x8 DCT matrix, set the hash value from 0 to 64 bits, set the hash value greater than or equal to the DCT mean value to 1, and set the hash value less than the DCT mean value to 0; S405, using the Hamming distance to calculate the difference hash similarity between the two images, and determine whether the target is a known target.

[0013] In one embodiment of the present invention, S50 includes: S501, dividing the pixels of the unknown target image into different regions, and treating all the regional data as a single cluster.

[0014] S502. Select a split data, usually the point with the largest variance within the cluster.

[0015] S503: Split the current cluster into multiple sub-clusters.

[0016] S504: Add the sub-clusters obtained by splitting to the cluster set.

[0017] S505. Repeat steps S502-S504 until each data is a separate cluster or a preset condition is met.

[0018] In a second aspect of the present invention, an electronic device is provided, including a hyperspectral image collector, an image processor, a computing processor and a memory, wherein the hyperspectral image collector is used to acquire images of a target area, the image processing module is used for image processing and classification, the computing processor is used to read the computing program, and the memory is used to store a standard image database and the above-mentioned program, and execute the target recognition method integrating hyperspectral remote sensing and machine learning as described in the first aspect.

[0019] In a third aspect of the present invention, a computer-readable storage medium is provided, on which instructions are stored. When the instructions are executed on a computer, the target recognition method integrating hyperspectral remote sensing and machine learning as described in the first aspect is executed.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. The present invention realizes the recognition of various different targets. After adding standard images of various known targets to the standard image database, the recognition of various targets is realized by extracting shape features and contour features in the standard images.

[0021] 2. Rapid target recognition is achieved through preprocessing, classification, feature extraction, and feature comparison of hyperspectral remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 A schematic diagram of the execution principle of a target recognition method based on the fusion of hyperspectral remote sensing and machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The above and other technical contents, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings. Through the description of the specific implementation mode, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and description, and are not used to limit the technical solution of the present invention.

[0025] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants are intended to cover non-exclusive inclusion, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "including one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0026] Example: In the first aspect, the present invention provides a target recognition method based on the fusion of hyperspectral remote sensing and machine learning. The establishment of a standard image database avoids the problem of slow speed in the early stage of machine learning and improves the efficiency of machine learning. After the hyperspectral remote sensing image is subjected to noise reduction and dimensionality reduction processing, the amount of calculation in the feature extraction process can be reduced and the efficiency of the feature can be accelerated. In feature extraction, the BRISK algorithm is used, which has stronger stability in the extraction of image features and reduces the error rate in target recognition. The PHash algorithm with relatively high accuracy is used in feature image comparison to ensure the accuracy of target recognition. The above-mentioned unknown target features are analyzed by a hierarchical clustering analysis algorithm and entered into the database to complete the learning.

[0027] Based on the above technical ideas, the execution principle of this method is as follows Figure 1 Specifically, the method comprises the following steps: S00, establishing a standard image database; the standard image database contains multiple feature images of multiple reference targets; each feature image contains shape features and contour features of the reference target; The standard image database contains multiple feature images of multiple reference targets, and also includes the pre-extraction of shape features and contour features in the target feature images, and the above features are stored in the standard image database to speed up the feature comparison.

[0028] S10, obtaining a hyperspectral remote sensing image of the target area through an image acquisition module; Specifically, an unmanned aerial vehicle or other platform is used to carry the device of the present invention, and a hyperspectral remote sensing image of a target area is acquired through an image acquisition module.

[0029] S20, performing noise reduction processing on each hyperspectral remote sensing image, performing dimensionality reduction processing on each hyperspectral remote sensing image after the noise reduction processing; performing image classification on the multiple hyperspectral remote sensing images after the dimensionality reduction processing; Specifically, the obtained hyperspectral remote sensing image is smoothed by the sparse smoothing preprocessing method, and then the targets in the image are classified so as to extract the target features in the next step. The specific steps are as follows: S201, decomposing the image into high-frequency components and low-frequency components through a global sparse regularization term; S202, selecting sample values ​​of smoothing parameters from high-frequency components by block division, then fitting a smoothing parameter surface using a Bessel method, and then calculating corresponding smoothing parameter values ​​according to the coordinates of the pixel points; S203, based on the theory that high-frequency components and low-frequency components are highly correlated, the low-frequency image is smoothed using a smoothing parameter obtained from the high-frequency component, and an image to be classified is obtained by iterative optimization; S204: classify different targets in the above-mentioned image to be classified according to preset standard values.

[0030] S30, extracting features from each type of hyperspectral remote sensing image to obtain multiple target feature images corresponding to each type of hyperspectral remote sensing image; each target feature image contains shape features and contour features of the target to be identified. Specifically, based on the BRISK algorithm, the target features of the classified hyperspectral images are extracted to generate the target feature image to be identified. The features are composed of target shape, contour, position and other features. The specific steps are as follows: S301. In constructing the feature descriptor, the BRISK algorithm uses a simple pixel grayscale value comparison to obtain a cascaded binary bit string to describe each feature point. S302, with the feature point as the center, construct multiple discretized Bresenham concentric circles of different radii, and then obtain N sampling points with the same spacing on each concentric circle. Since this neighborhood sampling mode will produce image grayscale aliasing during sampling, the standard deviation of the Gaussian function used is proportional to the spacing between points on each concentric circle.

[0031] S303, then rotating the sampling pattern around the feature point by a specified angle, so that the feature descriptor has rotation invariance.

[0032] S304: compare the grayscale values ​​of all feature point rows in the rotated short-distance sampling point set, and finally form a 512-bit binary string descriptor.

[0033] S305. Select feature points with larger RGB value difference, brightness and contrast from the descriptor obtained in the above manner, introduce directional gradient features and local structural features of the feature points, generate a binary comparison template, and calculate features of the area around the feature points, wherein the feature points with larger RGB value difference, brightness and contrast are feature points whose RGB value difference is higher than a preset threshold of the RGB value difference, whose brightness is higher than a preset threshold of the brightness, and whose contrast is higher than a preset threshold of the contrast.

[0034] On the basis of the above algorithm, a time series analysis algorithm is added to extract and record the target position features. Specifically, the position of the target is tracked by the changes of the basic particle filter at different times, and the motion trajectory of the target is modeled based on curve fitting.

[0035] S40, based on the PHash algorithm, compare and analyze the feature image generated in the previous step with the image in the preset standard database, and divide them into known targets and unknown targets; the specific steps are as follows: S401, PHash starts with a small image, which simplifies the calculation of DCT; S402, converting the image into a grayscale image to further simplify the amount of calculation; S403, calculating the DCT transform of the picture to obtain a 32x32 DCT coefficient matrix; S404, retain the 8x8 matrix in the upper left corner and calculate the mean value of DCT; S405, according to the 8x8 DCT matrix, set a hash value of 0 or 1 to 64 bits, set the hash value greater than or equal to the DCT mean value to 1, and set the DCT mean value to 0 when it rains; S406, using the Hamming distance to calculate the difference hash similarity between the two images, and based on this, determine whether the target is known.

[0036] S50, analyzing the above unknown target features through a hierarchical clustering analysis algorithm, and entering the features into a database to complete the learning.

[0037] Divisive hierarchical clustering is a top-down strategy that starts with all data points as a cluster and gradually splits the clusters until each data point is a separate cluster. In the image segmentation task, hierarchical clustering can help divide the pixels in the image into different regions. The main steps are as follows: S501, dividing the pixels of the unknown target image into different regions, and treating all the regional data as a single cluster.

[0038] S502. Select a split data, usually the point with the largest variance within the cluster.

[0039] S503: Split the current cluster into multiple sub-clusters.

[0040] S504: Add the sub-clusters obtained by splitting to the cluster set.

[0041] S505. Repeat steps S502-S504 until each data is a separate cluster or the preset condition is met.

[0042] S506: Complete the above analysis steps and enter the data into the database to complete machine learning.

[0043] Based on the method provided in the first aspect and the system provided in the second aspect, the second aspect of this embodiment provides an electronic device for executing the method described in the first aspect or any electronic device that may be involved in the method described in the first aspect, including a hyperspectral image collector, an image processor, a computing processor and a memory, wherein the hyperspectral image collector is used to obtain the target area image, the image processing module is used for image processing and classification, the computing processor is used to read the computing program, and the memory is used to store the standard image database and the above program. Specifically, the memory may include, but is not limited to, random access memory (Random-Access Memory, RAM), read-only memory (Read-Only Memory, ROM), flash memory (Flash Memory), first-input first-out memory (FirstInputFirstOutput, FIFO) and / or first-input last-out memory (FirstInputLastOutput, FILO), etc.; the processor may include, but is not limited to, a microprocessor of the STM32F105 series. In addition, the computer device may also include, but is not limited to, a power module, a display screen and other necessary components.

[0044] The working process, working details and technical effects of the aforementioned computer device provided in the second aspect of this embodiment can be referred to the method described in the first aspect or any method that may be related to the first aspect, and will not be described in detail here.

[0045] The third aspect of this embodiment provides a computer-readable storage medium that stores the method described in the first aspect or any method that may be related to the method described in the first aspect, that is, the computer-readable storage medium stores instructions, and when the instructions are run on a computer, the method described in the first aspect or any method that may be related to the method described in the first aspect is executed. The computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to computer-readable storage media such as floppy disks, optical disks, hard disks, flash memories, USB flash drives, and / or memory sticks (MemoryStick), and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0046] The working process, working details and technical effects of the aforementioned computer-readable storage medium provided in the third aspect of this embodiment can be referred to the method described in the first aspect or any method that may be related to the method described in the first aspect, and will not be repeated here.

[0047] In a fourth aspect, the present embodiment provides a computer program product comprising instructions, which, when executed on a computer, causes the computer to execute the method described in the first aspect or any method that may be related to the method described in the first aspect. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0048] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A target recognition method based on the fusion of hyperspectral remote sensing and machine learning, characterized in that: The following steps are involved: Establishing a standard image database; the standard image database contains multiple feature images of multiple reference targets; each feature image contains shape features of the reference target; Obtain multiple hyperspectral remote sensing images of the target area; Perform noise reduction processing on each hyperspectral remote sensing image, and perform dimension reduction processing on each hyperspectral remote sensing image after noise reduction processing; Perform image classification on multiple hyperspectral remote sensing images after dimensionality reduction processing; Perform feature extraction on each type of hyperspectral remote sensing image to obtain multiple target feature images corresponding to each type of hyperspectral remote sensing image; Each target feature image contains the shape features of the target to be identified; Based on the PHhash algorithm, the shape features of each target feature image are compared and analyzed with the shape features of each feature image to obtain the analysis results; The analysis results are processed using a hierarchical clustering analysis algorithm to output a target recognition result; the target recognition result includes known targets and unknown targets.

2. According to claim 1, a target recognition method based on the fusion of hyperspectral remote sensing and machine learning is characterized in that: The method for denoising each hyperspectral remote sensing image is: processing each hyperspectral remote sensing image based on a coefficient smoothing algorithm; specifically, the following steps are included: The spectral remote sensing image is decomposed into high-frequency components and low-frequency components through a global sparse regularization term; The high-frequency component is divided into blocks to select the sample value of the smoothing parameter, the smoothing parameter surface is fitted using the Bessel method, and the smoothing parameter value is obtained according to the coordinates of the pixel points; The low-frequency image is smoothed using the smoothing parameters obtained from the high-frequency component, and the image to be classified is obtained through iterative optimization. Different targets in the classified image are classified according to preset standard values.

3. According to claim 2, a target recognition method based on the fusion of hyperspectral remote sensing and machine learning is characterized in that: Based on the PHhash algorithm, the shape features of each target feature image are compared and analyzed with the shape features of each feature image, including the following steps: Use PHash algorithm to select pictures; Convert the selected image into a grayscale image; Calculate the DCT transform of the image and get a 32x32 DCT coefficient matrix; Keep the 8x8 matrix in the upper left corner of the image and calculate the mean of DCT; According to the DCT matrix, assign values ​​to the selected picture; The difference hash similarity between two images is calculated using the Hamming distance.

4. According to claim 3, a target recognition method based on the fusion of hyperspectral remote sensing and machine learning is characterized in that: The analysis results are processed using a hierarchical clustering analysis algorithm, including: Divide the pixels of the unknown target image into different regions. Select a split data; Split the current cluster into multiple sub-clusters; Add the sub-clusters obtained by splitting to the cluster set.

Citation Information

Patent Citations

  • Quick target identifying method and system based on compressed sensing

    CN106557784A

  • A method of substance recognition based on multi-spectra

    CN109214302A

  • Hyperspectral image classification experiment method

    CN112800827A

  • Hyperspectral image classification device and method

    CN116403046A

  • Multi-mode pasture image feature extraction and recognition system

    CN117197656A