CN-122023386-A - Multi-view agricultural image clustering method based on multi-order bipartite graph learning

CN122023386ACN 122023386 ACN122023386 ACN 122023386ACN-122023386-A

Abstract

The invention discloses a multi-view agricultural image clustering method based on multi-order bipartite graph learning, which relates to the technical field of image clustering, the method comprises the steps of optimizing an initial bipartite graph matrix by learning the direct relation between data points and anchor points through a first-order similarity learning module, extracting local features of a multi-view agricultural image, and taking a model as a first-order similarity learning model. And constructing a reconstruction error through a second-order similarity learning module, capturing an initial bipartite graph matrix of neighborhood structural feature optimization between the data points and the anchor points, and constructing a second-order similarity learning model by combining the first-order similarity learning module. Tensor is built through the enhanced third-order similarity learning module, the tensor Schatten-p norm and anchor point structure regularization are combined, high-order correlation of initial bipartite graphs among different view angles is mined, the relation between anchor points is considered, the first-order similarity learning module and the second-order similarity learning module are integrated, so that an enhanced multi-order similarity learning model is built, and the model is optimized to generate high-quality global representation.

Inventors

DENG YANGJUN
DENG WENHAO
REN LONGFEI
WANG WEIYE
Long Chenfeng
ZHU XINGHUI
LIU YANJIN
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: 20260319

Claims (8)

1. A multi-view agricultural image clustering method based on multi-order bipartite graph learning is characterized by comprising the following steps: Step S1, pre-selecting a representative sample as an anchor point, calculating a characteristic bipartite graph between a direct similarity construction data point and the anchor point through the data characteristic and the anchor point characteristic, constructing an initial bipartite graph which needs to be optimized and is used for representing the affinity between the data point and the anchor point, wherein the initial bipartite graph is a zero matrix, the optimization aim is to minimize the distance weighted sum between the data point and the anchor point according to the characteristic bipartite graph, meanwhile, the sum of similarity weights of all the anchor points corresponding to each data point is ensured to be one, and each weight value is between zero and one, so that a first-order similarity learning module between the data point and the anchor point is constructed, and the local characteristics of images with different view angles are respectively captured based on the first-order similarity learning module, and the interference of cloud and fog shielding and low-resolution noise on the local characteristics is relieved; Step S2, taking the initial bipartite graph in the step S1 as input, further analyzing the relation between the data points and the whole anchor point set, wherein the core is to optimize the initial bipartite graph in the step S1 by taking the reconstruction error of the weighted representation of the data points and the anchor points as a weighted value, and also follow the constraint condition that the sum of the weights of the anchor points corresponding to each data point is one and the weight value is between zero and one, so as to construct a second-order similarity learning module between the data points and the anchor points; And S3, constructing tensor Schatten-p norms and anchor point structure regularized enhanced third-order similarity learning, combining the first-order similarity learning module of the step S1 and the second-order similarity learning module of the step S2, wherein the third-order similarity learning module firstly constructs an initial bipartite graph matrix of each view as a third-order tensor, carries out constraint through tensor Schatten-p norms and anchor point structure regularization, finally fuses the optimized bipartite graphs of all the views through a cascading method to generate a consensus bipartite graph matrix, and then generates a final clustering result through spectral clustering aiming at the bipartite graphs and combining a k-means algorithm.
2. The multi-view agricultural image clustering method based on multi-order bipartite graph learning according to claim 1, wherein the step S1 includes the following steps: S11, preprocessing multi-view agricultural image data, extracting characteristic representation of each view angle, extracting a spectrum vector of each pixel point for a multi-spectrum image, generating texture characteristics corresponding to a space position for a visible light image, and extracting a temperature distribution vector for thermal imaging; s12, capturing local similarity characteristics between data points and anchor points and constructing an initial bipartite graph matrix Initializing to 0, optimizing the initial bipartite graph similarity weight by minimizing the Euclidean distance weighted sum between the data point and the anchor point to represent the first-order similarity, and constructing a first-order similarity learning model optimization objective function through a first-order similarity learning module, wherein the formula is as follows: ; In the formula, Represent the first The eigenvector of the data point, Represent the first The feature vectors of the individual anchor points, Representing the similarity weight of the data point to the anchor point, And The number of data points and anchor points respectively; In the process, the A negative value will not occur and, The weight is prevented from being too large; step S13, the optimization problem is equivalent to a minimum trace form, and the formula is as follows: In the formula, An edge weight matrix of the bipartite graph represents a direct relationship between original data points and anchor points, 。
3. The multi-view agricultural image clustering method based on multi-order bipartite graph learning according to claim 1, wherein the step S2 includes the following steps: Step S21, mining the neighborhood structure relation between the data points and the anchor point set according to the initial bipartite graph matrix generated in the step S1 as input, optimizing the initial bipartite graph similarity weight by minimizing the reconstruction error of the weighted representation of the data points and the anchor points so as to represent the second-order similarity, and optimizing the objective function, wherein the formula is as follows: step S22, the optimization target is equivalent to the minimum trace form, and the formula is as follows: In the formula, A data matrix, a feature vector containing all data points, For an anchor matrix, containing feature vectors for all anchors, tr (,) represents the matrix trace, and this optimization is accomplished by minimizing the data points Weighted representation with its anchor point Is ensured by the reconstruction error of (a) Reflecting the neighborhood structure similarity of the data points and the anchor point set; Step S23, constructing a second-order similarity learning model for simultaneously learning a global structure and a local structure by combining the objective function and the first-order similarity learning model between the data points and the anchor points in the step S1, wherein the global structure and the local structure complement each other, strengthen the clustering structure of the initial bipartite graph matrix, and inhibit the negative influence of noise points on the neighborhood characteristics, and the formula is as follows: 。
4. The multi-view agricultural image clustering method based on multi-order bipartite graph learning according to claim 1, wherein in the step S3, the obtained initial bipartite graph matrix of V views is obtained Stacking according to the view angle dimension, and performing rotation operation on the stacked matrix to obtain a third-order tensor with dimension of MXV XN M is the number of anchor points, N is the number of data points, and V is the total number of views.
5. The multi-view agricultural image clustering method based on multi-order bipartite graph learning according to claim 1, wherein the step S3 includes the following steps: Step S31, integrating the second-order similarity learning module, tensor Schatten-p norm constraint term and anchor point structure regularization term in step S2, thereby constructing an enhanced multi-order similarity learning model, realizing the mining of the high-order correlation between views on the basis of capturing the global structure and the local structure between data points and anchor points, further optimizing the clustering structure of the initial bipartite graph by considering the relation between the anchor points through the anchor point structure regularization, and the formula is as follows: In the formula, Represent the first A bipartite graph matrix of individual viewing angles, As a matrix of data, For the matrix of anchor points, For a bipartite graph edge weight matrix, a direct relationship between the original data point and the anchor point is represented, To construct a laplace matrix from a similarity matrix between anchor points, the construction process conforms to the general definition of a laplace matrix, For the tensor Schatten-p norm, Is a control parameter; S32, iterative optimization of an enhanced multi-order similarity learning model adopts an enhanced Lagrangian multiplier method, and the parameters are balanced 、 Controlling each loss weight and dynamically adjusting in the optimization process And finally, generating a consensus bipartite graph matrix by fusing bipartite graph matrices of different visual angles through cascade connection, obtaining an embedded vector by spectral clustering aiming at the bipartite graph matrix, and generating a final clustering result by combining a k-means algorithm.
6. The multi-view agricultural image clustering method based on multi-order bipartite graph learning according to claim 1, wherein the iterative optimization of step S3 and the computational complexity of the whole method are as follows Where T is the number of iterations of the augmented Lagrangian multiplier method, The number of data points is a number of data points, For the number of anchor points, Is the total number of views.
7. The multi-view agricultural image clustering method based on multi-order bipartite graph learning according to claim 1, wherein the feature preprocessing in step S1 includes a feature extraction operation and an optional operation: Extracting a spectrum vector of each pixel point from a multispectral image, generating a texture feature vector corresponding to a space position from a visible light image, extracting a temperature distribution vector from a thermal imaging image, and extracting a structure or texture feature vector of a corresponding scene from a high-resolution unmanned aerial vehicle aerial image; The method comprises the optional operation of denoising the image and normalizing pixel values aiming at the possible interferences such as sensor noise, cloud and fog shielding, illumination change and the like in the multi-view agricultural image so as to unify the characteristic numerical range or reduce the interference influence.
8. The multi-view agricultural image clustering method based on multi-level bipartite graph learning according to claim 1, wherein the method constructs a unified optimization framework in a progressive manner from a first-order similarity learning model to a second-order similarity learning model to an enhanced multi-level similarity learning model, and each module is mutually complemented and can be applied to multi-view agricultural image clustering scenes related to farmland zoning, crop classification, pest detection, crop health monitoring and yield prediction.

Description

Multi-view agricultural image clustering method based on multi-order bipartite graph 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 multi-order bipartite graph 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 shows explosive growth, and the multi-source image data cover visible light images (RGB), multispectral images, near infrared images, thermal imaging, high-resolution unmanned aerial vehicle aerial images and the like. The multi-view agricultural image data describes the characteristics of farmlands, crops and agricultural environments from different dimensions, and provides rich resources for accurate agriculture, crop health monitoring, pest detection and yield prediction. However, since images acquired by different sensors have heterogeneity including different spatial resolutions, spectral characteristics, temporal resolutions, and noise levels, it is difficult for conventional single-view clustering methods to effectively process these large-scale, complex data, and complementary information of multi-view data often cannot be fully mined, resulting in insufficient clustering accuracy. In the existing method, like a multi-view clustering technology based on subspace or spectral clustering, only single-order similarity modeling is usually focused, local features, neighborhood relations and high-order correlations are ignored, and complex distribution characteristics of agricultural images are difficult to comprehensively characterize. In addition, the traditional method has high calculation complexity when processing large-scale agricultural image data, is difficult to meet real-time analysis requirements, and has poor robustness to common interferences such as illumination changes, cloud and fog shielding, sensor noise and the like. Aiming at the problems, the invention provides a multi-view agricultural image clustering method based on multi-level bipartite graph learning, which is used for comprehensively capturing the local structure, neighborhood relation and higher-order correlation of multi-view agricultural image data points and anchor points and the relation between the anchor points by combining first-order, second-order and third-order similarities between the fused data points and the anchor points and combining tensor Schatten-p norm constraint and anchor point structure regularization. The method can effectively integrate complementary features of multispectral, visible light, thermal imaging and other data, construct stable similarity representation and remarkably improve clustering precision and robustness. Meanwhile, the computational complexity is reduced through an efficient optimization algorithm, so that the method is suitable for processing a large-scale agricultural image dataset. Compared with the traditional method, the method adopts a unified optimization framework in similarity learning and spectral clustering, avoids the problem of unstable performance caused by step-by-step optimization, is particularly suitable for processing complex agricultural scenes influenced by illumination, cloud and noise, and provides reliable and efficient clustering support for precise agricultural application. Disclosure of Invention In order to solve the technical problems, the invention is realized by the following technical scheme: The invention discloses a multi-view agricultural image clustering method based on multi-order bipartite graph learning, which comprises the following steps: Step S1, performing feature preprocessing on multi-view agricultural image data, wherein a spectrum vector of each pixel point is extracted for a multi-spectrum image, texture features of corresponding space positions are extracted for a visible light image, and a temperature distribution vector is extracted for a thermal imaging image; then preselecting a representative sample as an anchor point, constructing an initial bipartite graph between data points and the anchor point, wherein an initialization matrix is a zero matrix, optimizing the initial bipartite graph by means of a distance weighted sum (first-order similarity) between original data point features and original anchor point features, constructing a first-order similarity learning model by means of the first-order similarity learning module, simultaneously ensuring that the sum of similarity weights of all the anchor points corresponding to each data point is one, and each weight value is between zero and one, respectively capturing local features of images of different view angles by means of first-order similarity, and relieving interference of cloud shielding and low-resolution noise on the local features; step S2, taking an initial bipartite graph matrix