CN-122023384-A - Multi-view agricultural image clustering method based on enhanced multi-order similarity learning

CN122023384ACN 122023384 ACN122023384 ACN 122023384ACN-122023384-A

Abstract

The invention discloses a multi-view agricultural image clustering method based on enhanced multi-order similarity learning, which relates to the technical field of image clustering and comprises the following steps: preprocessing image data, constructing an initial affinity matrix for representing the affinity between data points, capturing the local structure and neighborhood relation of the data points through first-order similarity and second-order similarity, simultaneously, stacking all the affinity matrices to be rotated into third-order tensors, restraining by using weighted tensors Schatten-p norms, and optimizing all the modules in a unified optimization framework. According to the invention, the multi-order similarity, the weighted tensor Schatten-p norm optimization and the unified spectral clustering fusion are excavated in parallel, the complementary characteristics of multi-view data are fully utilized, the clustering precision and the robustness are improved, and reliable technical support can be provided for accurate agricultural applications such as crop health monitoring and pest detection.

Inventors

DENG YANGJUN
DENG WENHAO
LIU YANJIN
WANG XIONGWEI
REN LONGFEI
WANG WEIYE
ZHU XINGHUI
Long Chenfeng
SHEN LUMING

Assignees

湖南农业大学
长沙三昇农业科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260318

Claims (10)

1. The multi-view agricultural image clustering method based on the enhanced multi-order similarity learning is characterized by comprising the following steps of: step S1, first-order and second-order similarity learning, namely preprocessing multi-view agricultural image data to obtain a multi-view data matrix containing rich characteristic information, constructing an initial affinity matrix for the data matrix of each view, representing the affinity between data points, constructing a first-order similarity learning module, capturing local structural characteristics between the data points, constructing a second-order similarity learning module as a supplement of the first-order similarity learning module, and mining neighborhood structural relations between the data points, wherein constraint conditions of the two modules comprise constant c of the affinity matrix, symmetry of the affinity matrix, and 0-1 of element value of the affinity matrix; S2, constructing a third-order similarity learning module, namely obtaining third-order similarity through weighting tensor Schatten-p norm constraint, namely performing stacking and rotation operation on an initial affinity matrix to construct a third-order tensor reflecting high-order correlation among multiple views, combining a first-order similarity module, a second-order similarity module and a third-order similarity module on the basis, optimizing the initial affinity matrix, and adjusting a weighting vector Sum-power parameters The key information in the multi-view agricultural image is preferentially reserved, the noise related information is restrained, the weighted tensor Schatten-p norm optimization is realized through an objective function, and the objective function comprises balance parameters To control the influence of the weighted tensor Schatten-p norms, constraints ensure that third-order tensors are generated from the optimized affinity matrices for all views through stacking and rotation; And S3, unified spectral clustering fusion, namely introducing a spectral clustering constraint item constructed by a Laplacian matrix into the unified optimization framework based on the first-order similarity and second-order similarity optimization module in the step S1 and the third-order tensor optimization module in the step S2, wherein the Laplacian matrix is defined based on a fusion affinity matrix and a degree matrix, the fusion affinity matrix is calculated by the optimized affinity matrix of all view angles, the spectral clustering item comprises a clustering indication matrix F, a clustering structure is optimized by minimizing an objective function of the unified optimization framework, a final affinity graph with strong discriminant is generated, and the final affinity graph is subjected to spectral clustering to obtain a clustering result of the multi-view agricultural image.
2. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning of claim 1, wherein the multi-view agricultural image data comprises at least two of a visible light image, a multi-spectral image, a near infrared image, a thermal imaging image, and a high resolution unmanned aerial vehicle aerial image.
3. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning according to claim 1, wherein the module objective function for calculating the first-order similarity in step S1 is specifically: In the formula, Is the first An initial affinity matrix for each viewing angle; a similarity matrix for expressing the direct relationship between data points; for regularization parameters, controlling complexity of the affinity matrix to avoid invalid solutions; performing matrix trace operation; Is the Frobenius norm; And c is a constant.
4. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning according to claim 1, wherein the module objective function for calculating the second-order similarity in step S1 is specifically: In the formula, For the initial affinity matrix for the v-th viewing angle, The pre-processed initial data matrix for the v-th view, For the operation of the matrix trace, Is the Frobenius norm, V is the total number of views, c is a constant, Is an identity matrix.
5. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning of claim 1, wherein the specific process of constructing the third-order tensor in the step S2 is to perform initial affinity matrix Stacking according to the view angle dimension, and rotating the stacked matrix to obtain a dimension of Third order tensor of (2) Where N is the number of samples in the multi-view agricultural image data.
6. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning according to claim 1, wherein the laplace matrix in step S3 Is defined as: In the formula, In order to fuse the affinity matrix, Is that A corresponding degree matrix.
7. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning of claim 6, wherein the fusion affinity matrix The calculation mode of (a) is as follows: In the formula, For the initial affinity matrix for the V-th view, V is the total number of views.
8. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning according to claim 1, wherein the size of the clustering indication matrix F in the step S3 is , wherein, Is the number of clusters to be preset, Is the number of samples and the cluster indicates that matrix F satisfies the orthogonality constraint.
9. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning according to claim 1, wherein the preprocessing in step S1 includes image denoising, pixel value normalization, and feature extraction, and the features extracted by the feature extraction include at least one of crop morphology features, soil texture features, crop spectrum features, and temperature distribution features.
10. The multi-view agricultural image clustering method based on enhanced multi-order similarity learning according to claim 1, wherein in step S2, the weighting vector is optimized by means of experimental parameter adjustment Sum-power parameters The experimental parameter adjustment takes the accuracy of multi-view agricultural image clustering as an evaluation index, and selects the highest accuracy And As an optimal parameter.

Description

Multi-view agricultural image clustering method based on enhanced multi-order similarity learning Technical Field The invention belongs to the technical field of image clustering, and particularly relates to a multi-view agricultural image clustering method based on enhanced multi-order similarity learning. Background With the rapid development of remote sensing technology, unmanned aerial vehicle technology and multi-sensor imaging technology, multi-source image data acquired in the agricultural field show explosive growth. Such data includes, but is not limited to, visible light images (RGB), multispectral images, near infrared images, thermal imaging, high resolution unmanned aerial vehicle aerial images, and the like. The multi-source image data describe the characteristics of farmlands, crops and agricultural environments from different view angles and dimensions, and provide abundant resources for applications such as accurate agriculture, crop health monitoring, pest detection, yield prediction and the like. However, since image data acquired by different sensors or view angles have different spatial resolution, spectral characteristics, temporal resolution and noise characteristics, how to efficiently and cooperatively use these multi-source multi-view image data, and to mine complementary information thereof to realize high-precision agricultural image clustering, has become a research hotspot in the current agricultural informatization field. The multi-view agricultural image data contains rich complementary information. For example, multispectral images can provide spectral features of crops for distinguishing crop types or detecting health states, but the spatial resolution is generally low, and fine spatial structures are difficult to capture, while visible light images have higher spatial resolution, can clearly reflect the morphology and texture features of crops, but lack spectral information, and thermal imaging can reflect the temperature distribution of crops or soil, is helpful for detecting water stress or diseases, and is easily disturbed by environmental temperature. Image data of a single view is difficult to comprehensively represent complex characteristics of an agricultural scene due to limitations of the image data, and collaborative analysis of multi-view data can reduce noise influence of a single data source by fusing complementary characteristics of different views, so that clustering accuracy and robustness are improved. Therefore, how to effectively utilize the local features, neighborhood relations and higher-order correlations of the multi-view agricultural image data to construct stable similarity representation and realize efficient clustering is a core problem to be solved by the multi-view agricultural image clustering technology. In the prior art, the multi-view clustering method is mainly based on subspace clustering, spectral clustering or low-rank representation and other technologies, and clustering performance is improved by mining consistency and complementarity of multi-view data. However, the conventional method has the defects that firstly, most methods are insufficient in mining similarity characteristics (such as local structures, neighborhood relations or higher-order correlations) of different view angles, complex characteristics of agricultural images are difficult to comprehensively characterize, secondly, the conventional method generally adopts uniform regularization treatment on singular values of a similarity matrix, importance degree differences of the singular values are ignored, key information is lost, and in addition, similarity learning and spectral clustering processes are generally carried out step by step, and a uniform optimization framework is lacked, so that clustering performance is unstable. In particular, in an agricultural scene, image data is often influenced by factors such as illumination change, cloud and fog shielding, sensor noise and the like, and the robustness and adaptability of the traditional method in processing the complex data are poor, so the following scheme is proposed for solving the problems. Disclosure of Invention The invention aims to provide a multi-view agricultural image clustering method based on enhanced multi-order similarity learning, which can improve clustering robustness and precision by parallelly mining multi-order similarity, weighting tensor Schatten-p norm optimization and unified spectral cluster fusion and solves the problems of insufficient mining and poor robustness of the existing multi-view agricultural image clustering. In order to solve the technical problems, the invention is realized by the following technical scheme: the invention relates to a multi-view agricultural image clustering method based on enhanced multi-order similarity learning, which specifically comprises the following steps: Step S1, first-order and second-order similarity learning, namely preprocessing multi-view agric