An ornamental plant classification and identification method based on a graph neural network

Abstract

This invention discloses a method for classifying and recognizing ornamental plants based on graph neural networks, comprising the following steps: acquiring sample images of ornamental plants, performing preprocessing and organ region segmentation; calling the GSG network to construct a curvelet scattering module and performing multi-scale directional decomposition; mapping directional feature representations to edge space to construct a topological representation, preserving node connectivity and directional propagation relationships; performing Hodge decomposition on the topological representation, dividing it into three branches; constructing GSG scattering pyramids in each of the three branches, progressively expanding the depth and scale; performing feature alignment and weighted fusion to generate plant-level feature vectors; inputting the plant-level feature vectors into a classification module, outputting classification probabilities and category labels, and generating an organ region contribution heatmap. This invention enhances multi-scale feature extraction and orientation-sensitive modeling capabilities, improving the accuracy and stability of plant classification and recognition.

Description

Technical Field

[0001] This invention relates to the field of plant taxonomy and plant identification, and in particular to a method for classifying and identifying ornamental plants based on graph neural networks. Background Technology

[0002] The classification and identification of ornamental plants are of great significance in horticultural research, plant resource conservation, and industrial applications. Existing methods largely rely on manual observation and traditional image processing techniques, classifying plants by extracting features such as leaf morphology and petal texture. However, these methods generally suffer from limitations in feature simplification and robustness, making it difficult to adapt to the diverse morphological, scale, and growth environment variations among different plant varieties, resulting in insufficient stability and accuracy of classification results.

[0003] In recent years, with the development of deep learning, some studies have attempted to use convolutional neural networks (CNNs) for plant recognition, which has improved classification performance to some extent. However, CNNs are mainly based on image modeling using regular grids. For plants, which naturally possess complex topological relationships, the interrelationships between leaves, petals, and stems are difficult to capture effectively, resulting in limited feature representation capabilities. Furthermore, existing methods are still insufficient in modeling multi-scale and multi-directional structural details, often neglecting the interdependence between local organ features and the overall structure.

[0004] In addition, although some studies have attempted to introduce graph structures for representation, common methods have shortcomings in handling local directional information and topological consistency, and cannot effectively distinguish different types of structural components, resulting in feature redundancy or incomplete expression, which ultimately affects the generalization performance of classification models.

[0005] Therefore, how to provide a method for classifying and identifying ornamental plants based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for classifying and recognizing ornamental plants based on graph neural networks. This invention constructs an organ-level graph structure and combines multi-scale directional decomposition and direction indexing mechanisms of a curve scattering module to extract local features of plants at different scales and directions. Furthermore, it performs direction-sensitive Hodge decomposition in the edge space to distinguish between gradient flow, curl flow, and harmonic flow, constructs a multi-level GSG scattering pyramid, and fuses local directional features with global topological features to generate plant-level feature vectors and achieve classification. This invention balances fine-grained representation of organ structures with modeling of overall topological relationships, possessing advantages such as high recognition accuracy, strong feature interpretability, and superior generalization ability.

[0007] A method for classifying and identifying ornamental plants based on a graph neural network according to an embodiment of the present invention includes the following steps: Images of ornamental plant samples are acquired, image preprocessing is performed, organ regions are segmented by combining morphological features, organ-level graphs are constructed, and node features and edge attributes are generated in the graphs. Based on organ-level graphs, the GSG network is invoked to construct a curve scattering module, which performs multi-scale directional decomposition on node features, generates directional feature representations, and simultaneously generates directional indices. The directional feature representation is mapped to the edge space, and the topological representation is constructed by combining the edge attributes. The connection relationship and directional transmission relationship between nodes are preserved in the topological representation. Perform Hodge decomposition on the topological representation to divide the edge space features in the topological representation into gradient flow branch, curl flow branch and harmonic flow branch; In the gradient flow branch, the curl flow branch, and the harmonic flow branch, GSG scattering pyramids are constructed respectively, and the decomposition depth and scale are expanded layer by layer to extract multi-level flow feature representations; Based on directional feature representation and multi-level flow feature representation, feature alignment and weighted fusion are performed to generate plant-level feature vectors; The plant-level feature vector is input into the classification module, the feature classification operation is performed, the classification probability and category label are output, and the organ region contribution heatmap is generated by combining the directional index.

[0008] Optionally, the generation of the organ-level graph, node features, and edge attributes includes: Based on images of ornamental plant samples, image preprocessing operations are performed, including image denoising, geometric distortion correction, and resolution unification. Based on the preprocessed image, an organ region segmentation method is used to generate organ region segmentation results, extract the node set and edge set, and based on the node set and edge set, an organ-level graph is generated using a graph structure construction method. Based on the organ region segmentation results, morphological features are extracted, a morphological feature set is generated, and the morphological feature set is written into the node features. Based on organ-level graphs and node features, a geometric feature encoding method is used to map the distance, angular relationship, and curvature feature combination between organ nodes into edge attributes.

[0009] Optionally, the generation of the direction feature representation and direction index includes: Based on organ-level maps and node features, the GSG network is invoked to construct a curve scattering module and set multi-scale directional decomposition parameters to generate a set of decomposition parameters. Based on the curve wave scattering module and decomposition parameter set, the curve wave multi-scale directional decomposition method is used to decompose the node features, extract local structural features at different scales and directions, and generate directional feature representations. Based on directional feature representation, the scale parameters and directional parameters generated during the multi-scale directional decomposition of curve waves are written into the index structure to generate a directional index.

[0010] Optionally, the construction of the topological representation includes: Based on the directional feature representation and the edge set of the organ-level graph, the node features are mapped to the edge space to generate the initial edge space features; Based on the initial edge space features and edge attributes, the features are fused to generate fused edge space features; Based on the integration of edge space features, the connection relationships between nodes are constructed; Encode the direction index, pass the direction information to the edge space, and generate the direction transfer relationship; A topological representation is constructed based on the fusion of edge space features, node connection relationships, and direction propagation relationships.

[0011] Optionally, the generation of the three types of structured edge features includes: Based on topological and directional feature representations, edge space features are extracted and discrete exterior differential operators and edge Laplacian operators are constructed to form a set of decomposition operators; Based on the direction index, under the action of the decomposition operator set, the direction-sensitive Hodge decomposition is performed on the edge space features, and the edge space features are decomposed into gradient components, curl components and harmonic components under different directions respectively. Under the direction index, the edge space propagation mechanism is invoked to accumulate the gradient components in the edge set of the topological representation with direction weights. The local gradient values ​​are propagated and smoothed through the direction-related edge weights to generate gradient flow branches. The loop aggregation mechanism is invoked to perform directed cyclic accumulation of curl components in the loop set of topological representation. The rotation propagation intensity is calculated through the edge connection relationship within the loop, and feature transfer paths are established in the direction-related loops to generate curl flow branches. The global consistency mechanism is invoked to project the harmonic components onto the edge Laplacian operator null space in the topological representation. The global projection maintains the directional consistency across organ regions, and a normalization mapping is performed on the direction-dependent overall topological structure to generate harmonic flow branches.

[0012] Optionally, the generation of the multi-level flow feature representation includes: Based on gradient flow branch, curl flow branch and harmonic flow branch, the number of levels, scale range and number of directional channels of GSG scattering pyramid are set respectively, and scale transfer parameters between levels are established. The gradient component, curl component, and harmonic component are input into the GSG scattering pyramid, and the GSG scattering operation unit is called. In the first layer, the curvelet scattering transformation is performed to extract component features, and the first layer features are output by combining nonlinear activation and aggregation mechanisms. In the higher layers, the lower-level features are recursively input, and the scale transfer parameter is called to control the scale expansion relationship. The curvelet scattering transformation, nonlinear activation, and aggregation are continued until all layers are completed, and the multi-scale hierarchical features of the gradient flow branch, curl flow branch, and harmonic flow branch are obtained respectively. Based on multi-scale hierarchical features of gradient flow branch, curl flow branch and harmonic flow branch, channel alignment, scale alignment and direction index alignment are performed, weighted fusion unit is called, features are weighted and converged at the same level, and layer-by-layer merging is performed between levels to generate multi-level flow feature representations covering local and global aspects.

[0013] Optionally, the generation of the plant-level feature vector includes: Based on directional feature representation and multi-level flow feature representation, channel alignment, scale alignment and directional index alignment are performed to generate an aligned feature set; Based on the aligned feature set, the branch weights and scale weights are calculated using the orientation index to generate a fusion weight set; Based on the aligned feature set and the fusion weight set, different branches are weighted and superimposed at the same level, and layer-by-layer convergence is performed between levels to generate fused features; Based on the fusion features, global aggregation and vectorization mapping are performed to generate plant-level feature vectors.

[0014] Optionally, the generation of the classification probability, category label, and organ region contribution heatmap includes: Based on plant-level feature vectors, the classification module performs forward inference to generate classification probabilities and determine class labels, while backpropagation is performed to generate gradient information for the classification module. Based on the gradient information of the classification module, the gradient attribution method is used to calculate the contribution of plant-level feature vectors to the directional channel, and the directional contribution weight is generated by combining the directional index. Based on the directional contribution weight and the organ-level graph, the directional contribution weight is propagated along the organ node set and edge set, and then back-projected to the organ region according to the organ node to generate an organ region contribution heatmap.

[0015] The beneficial effects of this invention are: First, by constructing an organ-level graph and introducing node features and edge attributes into it, a refined representation of the local organ structure of ornamental plants is achieved, effectively avoiding the problem of insufficient capture of local features by traditional whole-image recognition methods. Secondly, by introducing a curved wave scattering module and a multi-scale directional decomposition mechanism with orientation index, rich structural features can be extracted at different scales and orientations, making the feature representation more robust and orientation sensitive. Furthermore, by performing direction-sensitive Hodge decomposition in the edge space and combining it with three branches—gradient flow, curl flow, and harmonic flow—the ability to model complex topological relationships is further enhanced. Finally, based on the multi-level GSG scattering pyramid and feature alignment fusion mechanism, local directional features and global topological features are unified into plant-level feature vectors, ensuring the accuracy and stability of classification, thereby achieving high-precision and interpretable ornamental plant classification and recognition results. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an ornamental plant classification and identification method based on graph neural networks proposed in this invention. Figure 2 This is a schematic diagram of the branch structure based on direction-sensitive Hodge decomposition proposed in this invention; Figure 3 This is a schematic diagram of the multi-level feature extraction structure based on the GSG scattering pyramid proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1-3 A method for classifying and recognizing ornamental plants based on graph neural networks includes the following steps: Images of ornamental plant samples are acquired, and image preprocessing operations are performed, including image denoising, geometric distortion correction and resolution unification. Organ regions are segmented by combining morphological features, and an organ-level graph containing a set of nodes and a set of edges is constructed. Node features and edge attributes are generated in the graph, and leaf length, leaf width, petal shape and stem curvature are written into the node features. Based on organ-level graphs, the GSG network is invoked to construct a curve scattering module, which performs multi-scale directional decomposition on node features, extracts local structural information at different scales and directions, generates directional feature representations, and simultaneously generates directional indexes to record the scale and direction parameters of feature decomposition. The directional feature representation is mapped to the edge space, and the topological representation is constructed by combining the edge attributes. The connection relationship and directional transmission relationship between nodes are preserved in the topological representation, providing input for branch decomposition. Perform Hodge decomposition on the topological representation to divide the edge space features in the topological representation into gradient flow branch, curl flow branch and harmonic flow branch, which respectively represent the local gradient component, loop rotation component and global harmonic component, resulting in three types of structured edge features; In the gradient flow branch, the curl flow branch, and the harmonic flow branch, GSG scattering pyramids are constructed respectively. The decomposition depth and scale are expanded layer by layer to extract multi-level flow feature representations and enhance the ability to characterize plant structures at different topological scales. Based on directional feature representation and multi-level flow feature representation, feature alignment and weighted fusion are performed to generate plant-level feature vectors, ensuring that local directional features and global topological features are represented in a unified vector space. The plant-level feature vector is input into the classification module, the feature classification operation is performed, the classification probability and category label are output, and the organ region contribution heatmap is generated by combining the directional index.

[0019] In this embodiment, the generation of the organ-level graph, node features, and edge attributes includes: Based on images of ornamental plant samples, image preprocessing operations are performed, including image denoising, geometric distortion correction, and resolution unification. Based on the preprocessed image, an organ region segmentation method is used to generate organ region segmentation results, extract the node set and edge set, and based on the node set and edge set, an organ-level graph is generated using a graph structure construction method. Based on the organ region segmentation results, morphological features are extracted to generate a set of morphological features including leaf length, leaf width, petal shape features and stem curvature. The set of morphological features is written into the node features. Leaf length and leaf width are calculated based on the main axis and secondary axis of the leaf outline. Petal shape features are described based on the ratio of the major axis to the minor axis of the petal outline and the concavity and convexity of the boundary. Stem curvature is calculated based on the degree of curvature of the stem centerline. Based on organ-level graphs and node features, a geometric feature encoding method is used to map the distance, angular relationship, and curvature features between organ nodes into edge attributes. The edge attributes include the Euclidean distance between nodes, angular relationship, and connectivity between adjacent organs. Output organ-level graph, node features, and edge attributes.

[0020] In this embodiment, the generation of the direction feature representation and direction index includes: Based on organ-level maps and node features, the GSG network is invoked to construct a curve scattering module and set multi-scale directional decomposition parameters to generate a set of decomposition parameters. Based on the curve wave scattering module and decomposition parameter set, the curve wave multi-scale directional decomposition method is used to decompose the node features, extract local structural features at different scales and directions, and generate directional feature representations. Specifically, firstly, a set of decomposition parameters is loaded into the curve scattering module to map the node features in the organ-level map to the spectral domain space. Then, the overall contour features of the organ region are extracted at the low-frequency scale, the morphological edge features are extracted at the mid-frequency scale, and the fine-grained texture features are extracted at the high-frequency scale. Within each scale, directional response calculation is performed based on a family of filters with multiple directional angles to capture the variation patterns of node features in different directions. Finally, the response results at different scales and directions are reprojected back into the node space, encoded as local structural features, and combined to generate directional feature representations. Based on directional feature representation, the scale parameters and directional parameters generated during the multi-scale directional decomposition of curve waves are written into the index structure to generate a directional index. The output direction feature representation and direction index are used as input for subsequent topology representation construction.

[0021] In this embodiment, the construction of the topological representation includes: Based on the directional feature representation and the edge set of the organ-level graph, the node features are mapped to the edge space to generate the initial edge space features; Based on the initial edge space features and edge attributes, the features are fused to generate fused edge space features; Based on the fusion of edge space features, the connection relationships between nodes are constructed to represent the topological structure between organs; Encode the direction index, pass the direction information to the edge space, and generate the direction transfer relationship; A topological representation is constructed based on the fusion of edge space features, node connection relationships, and directional transmission relationships, while simultaneously preserving edge space features, connection relationships, and directional transmission relationships in the topological representation.

[0022] In this embodiment, the generation of the three types of structured edge features includes: Based on topological representation and directional feature representation, edge space features in the topological representation are invoked and discrete exterior differential operators and edge Laplace operators are constructed to form a set of decomposition operators; Based on the direction index, under the action of the decomposition operator set, the direction-sensitive Hodge decomposition is performed on the edge space features, and the edge space features are decomposed into gradient components, curl components and harmonic components under different directions respectively. The specific process of the direction-sensitive Hodge decomposition is as follows: The current direction channel is determined one by one according to the direction index; the node response of that direction is extracted from the direction feature representation and aligned with the topological representation; the node response is mapped to direction-related edge space features; the discrete exterior differential operator and the edge Laplacian operator are called in the decomposition operator set to sequentially extract components from the direction-related edge space features. First, the gradient component is extracted through the gradient operator to represent the direction change along the edge; second, the curl component is extracted through the curl operator to represent the rotation propagation within the loop; then, the harmonic component is extracted by projection into the null space of the edge Laplacian operator to represent the global consistency relationship; the three types of components are scaled and normalized, and the direction index is bound to form direction-sensitive gradient components, direction-sensitive curl components, and direction-sensitive harmonic components, which are then output for subsequent generation of gradient flow branches, curl flow branches, and harmonic flow branches. The gradient operator is a sidespace operator constructed from discrete exterior differential operators. Its function is to map the node response to the sidespace and generate gradient components that can characterize local directional changes in the organ-level graph by calculating the numerical differences between adjacent nodes under the directional feature representation. The curl operator is a sidespace operator jointly constructed by the discrete external differential operator and the loop structure in the topological representation. Its function is to further map the sidespace features to the loop set, and generate a curl component that can characterize the loop rotation propagation characteristics of the organ region in the organ-level map by calculating the directed cumulative difference of the side features within the loop. Under the direction index, the edge space propagation mechanism is invoked to accumulate the gradient components in the edge set of the topological representation with direction weights. The local gradient values ​​are propagated and smoothed through the direction-related edge weights to generate gradient flow branches, which are used to characterize the local change relationship of the organ region in different directions. The loop aggregation mechanism is invoked to perform directed cyclic accumulation of curl components in the topologically represented set of loops. The rotational propagation intensity is calculated through the edge connection relationship within the loop, and feature transfer paths are established in the direction-related loops to generate curl flow branches, which are used to characterize the rotational propagation relationship of organ region loops in different directions. The rotational propagation intensity represents the cumulative amplitude of the curl component in a specific direction within the loop set. It is used to measure the intensity of direction-dependent rotational effects in the loop, and its value is determined by the internal edge connectivity and direction index of the loop, reflecting the distribution of loop rotational energy in different directions of the organ region. The rotational propagation intensity is input as a direction-dependent weighting factor into the loop aggregation mechanism to adjust the propagation weight of the curl component in the feature transmission path, thereby preserving direction-sensitive rotational propagation features in the generated curl flow branches. The global consistency mechanism is invoked to project the harmonic components onto the edge Laplacian operator null space in the topological representation. The global projection maintains the directional consistency across organ regions, and a normalization mapping is performed on the direction-related overall topological structure to generate harmonic flow branches, which are used to characterize the global consistency relationship of the overall organ structure in different directions.

[0023] In this embodiment, the generation of the multi-level flow feature representation includes: Based on gradient flow branch, curl flow branch and harmonic flow branch, the number of levels, scale range and number of directional channels of GSG scattering pyramid are set respectively, and scale transfer parameters between levels are established to control the depth and directional resolution of layer-by-layer expansion. The number of levels is set according to the size and structural complexity of the node set and edge set in the organ-level graph, the scale range is set according to the spatial resolution of the input image and the morphological size distribution of the plant organs, and the number of directional channels is set according to the resolution of the directional parameters recorded in the directional index, so as to ensure that the decomposition process can simultaneously cover local details and global structure and be consistent with the directional feature representation. The establishment of the scale transfer parameters between the layers is specifically as follows: Based on the scale range and the number of layers, the scale sequence is divided according to the scale step size of the curve scattering module; directional channels are assigned to each scale according to the directional index to form a scale-direction mapping; adjacency expansion radius and edge weight scaling factor are set for each layer on the organ-level graph to determine the feature transfer ratio from lower to higher layers; a downsampling factor and filtering bandwidth are defined for each layer, and the center frequency and support domain of the curve scattering filter are limited; an inter-layer connection matrix is ​​constructed to specify the splicing order and aggregation method of the output of the previous layer into the input of the next layer; cross-layer normalization coefficients and numerical stability thresholds are set simultaneously to ensure the comparability of feature amplitudes at different scales; finally, a scale transfer parameter set containing the scale sequence, scale-direction mapping, downsampling factor, filtering bandwidth, adjacency expansion radius, edge weight scaling factor, and inter-layer connection matrix is ​​generated. The gradient component, curl component, and harmonic component are input into the GSG scattering pyramid, and the GSG scattering operation unit is called. In the first layer, the curvelet scattering transformation is performed to extract component features, and the first layer features are output by combining nonlinear activation and aggregation mechanisms. In the higher layers, the lower-level features are recursively input, and the scale transfer parameter is called to control the scale expansion relationship. The curvelet scattering transformation, nonlinear activation, and aggregation are continued until all layers are completed, and the multi-scale hierarchical features of the gradient flow branch, curl flow branch, and harmonic flow branch are obtained respectively. The scaling parameter control method for scaling relationships includes: calling the curve filter layer by layer based on the scaling sequence to determine the scaling range and center frequency of the current layer; assigning a directional index to each scale based on the scaling-direction mapping to ensure that the directional response and directional feature representation are consistent across different scales; downsampling the low-level features in the hierarchical recursion based on the downsampling factor and extracting a stable frequency domain response under the filter bandwidth limit; expanding the local adjacency range on the edge set of the organ-level graph by calling the adjacency expansion radius and adjusting the weight strength of the edge attributes in combination with the edge weight scaling coefficient to achieve layer-by-layer propagation of the multi-level edge space; calling the inter-layer connection matrix to specify the splicing order and aggregation method of the output of the previous layer into the input of the next layer to ensure the structural consistency of the multi-scale hierarchical features; introducing a normalization coefficient in the cross-layer feature merging process to ensure that the amplitudes of features at different scales are comparable, and suppressing abnormal amplification or attenuation under the numerical stability threshold constraint, ultimately forming a scaling process jointly driven by the scaling sequence, scaling-direction mapping, downsampling factor, filter bandwidth, adjacency expansion radius, edge weight scaling coefficient, and inter-layer connection matrix. In the gradient flow branch, the GSG scattering operation unit is called to input the gradient components into the first layer of the scattering pyramid, and the curvelet scattering transformation is performed to extract local gradient direction features. Then, nonlinear activation and local aggregation are performed to output the first layer gradient features. Subsequently, low-level features are recursively input into higher levels, and scattering transformation, nonlinear activation and local aggregation are performed again until all levels are completed, resulting in multi-scale hierarchical features of the gradient flow branch. In the curl flow branch, the GSG scattering operation unit is called to input the curl component into the first layer of the scattering pyramid. Curved wave scattering transformation is performed to extract the loop rotation features. Then, nonlinear activation and local aggregation within the loop are performed to output the first layer of curl features. Subsequently, low-level features are recursively input into higher levels, and scattering transformation, nonlinear activation and aggregation are performed again until all levels are completed, resulting in the multi-scale hierarchical features of the curl flow branch. In the harmonic flow branch, the GSG scattering operation unit is called to input the harmonic components into the first layer of the scattering pyramid. Curved wave scattering transformation is performed to extract globally consistent features. Then, nonlinear activation and global aggregation are performed to output the first layer of harmonic features. Subsequently, low-level features are recursively input into higher levels, and scattering transformation, nonlinear activation and aggregation are performed again until all levels are completed, resulting in the multi-scale hierarchical features of the harmonic flow branch. Based on multi-scale hierarchical features of gradient flow branch, curl flow branch and harmonic flow branch, channel alignment, scale alignment and direction index alignment are performed, weighted fusion unit is called, features are weighted and converged at the same level, and layer-by-layer merging is performed between levels to generate multi-level flow feature representations covering local and global aspects. The output is a multi-level flow feature representation, which serves as the input for subsequent feature alignment and weighted fusion steps.

[0024] In this embodiment, the generation of the plant-level feature vector includes: Based on directional feature representation and multi-level flow feature representation, channel alignment, scale alignment and directional index alignment are performed to generate an aligned feature set; Based on the aligned feature set, the branch weights and scale weights are calculated using the orientation index to generate a fusion weight set; Based on the aligned feature set and the fusion weight set, a weighted fusion method is used to superimpose different branches at the same level, and layer-by-layer aggregation is performed between levels to generate fused features; Based on the fusion features, global aggregation and vectorized mapping are performed to generate plant-level feature vectors; Output plant-level feature vectors as input to the classification module.

[0025] In this embodiment, the generation of classification probability, category label, and organ region contribution heatmap includes: Based on plant-level feature vectors, the classification module performs forward inference to generate classification probabilities and determine class labels, while backpropagation is performed to generate gradient information for the classification module. The classification module includes an input layer, a fully connected layer, and a Softmax function. The input layer receives plant-level feature vectors and maps them to a high-dimensional representation. The fully connected layer performs a linear transformation on the high-dimensional representation to generate a category score vector. The Softmax function performs a normalization operation on the category score vector, mapping the category scores to a classification probability distribution. During the forward inference process, the plant-level feature vectors pass through the input layer, the fully connected layer, and the Softmax function in sequence, finally outputting the classification probability. The label corresponding to the category with the highest probability value is selected as the category label. At the same time, the partial derivative of the classification probability with respect to the plant-level feature vector is obtained to obtain gradient information. The gradient information is the parameter sensitivity result generated by the classification module during backpropagation. Its generation process includes: first, calculating the partial derivative of the classification probability with respect to the plant-level feature vector to obtain the gradient value of the classification probability as a function of the input; then, backpropagating the gradient value along the weight matrix of the fully connected layer of the classification module to calculate the influence of each input dimension on the class score; finally, combining the backpropagation result with the probability distribution of the Softmax function to map it into a sensitivity index on the directional channels, thereby obtaining gradient information characterizing the rate of change of the classification probability in different directional channels, used to quantify the contribution of each directional channel of the plant-level feature vector to the classification result; Based on the gradient information of the classification module, the gradient attribution method is used to calculate the contribution of plant-level feature vectors to the directional channel, and the directional contribution weight is generated by combining the directional index. The gradient attribution method binds the gradient information of the classification module with the orientation index to establish a mapping relationship between the orientation channel and the contribution. The orientation index is used to identify the scale parameters and orientation parameters generated by the curve multi-scale directional decomposition, and the gradient information is used to quantify the sensitivity of the classification probability to each orientation channel. The orientation contribution weight is generated by the joint calculation of the orientation index and the gradient information, which is used to characterize the classification contribution of the plant-level feature vector in different orientation channels. Based on the directional contribution weight and organ-level graph, the directional contribution weight is propagated along the organ node set and edge set, and back-projected to the organ region according to the organ node to generate the organ region contribution heatmap. Output classification probability, category label, and organ region contribution heatmap.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to the classification and identification of ornamental plants, and a complete experimental procedure and data verification system were constructed. Sample images of various ornamental plants were selected within the scenario, encompassing diverse features such as large leaves, numerous petals, complex flower colors, and significant stem curvature, to test the accuracy and stability of the classification method under complex backgrounds and various growth forms. The core of the experiment was to verify whether the structured representation method based on graph neural networks could more comprehensively capture organ-level information of plants, and to significantly improve the classification and identification effect by extracting multi-level features through direction-sensitive Hodge decomposition and GSG scattering pyramid.

[0027] In practical applications, input data is acquired by a high-definition image acquisition device, with the resolution uniformly adjusted to 1024×1024 pixels to ensure data consistency among different plant samples. The images undergo preprocessing steps, including denoising, geometric distortion correction, and resolution unification. Subsequently, organ region segmentation methods are used to clearly distinguish leaf, petal, and stem regions, constructing organ-level maps. In the organ-level maps, node features not only include geometric dimension data but also key morphological features such as leaf length, leaf width, petal shape, and stem curvature. Edge attributes consist of the distance, angle, and connectivity between organs.

[0028] The method of this invention feeds these structured inputs into a GSG network, utilizes curvelet multi-scale directional decomposition to capture local feature changes at different scales and directions, and then uses direction-sensitive Hodge decomposition to distinguish gradient components, curl components, and harmonic components, thereby modeling local organ changes, loop rotation propagation, and global consistency relationships. Subsequently, through recursive expansion of the GSG scattering pyramid at different levels, multi-level directional and topological features are extracted. Finally, feature alignment and weighted fusion are used to generate a plant-level feature vector. The classification module receives this vector and outputs the classification probability and class label, while simultaneously combining the directional index to generate an organ region contribution heatmap, thus ensuring the interpretability of the classification results.

[0029] To quantify the beneficial effects of this invention, comparative experiments were conducted with traditional convolutional neural network (CNN) methods and conventional graph convolutional network (GCN) methods. The experimental data comprised 500,000 images of ornamental plants, covering 300 plant classes, with a balanced number of samples per class. The ratio of the test set to the training set was set to 2:8. Evaluation metrics included classification accuracy, precision, recall, F1 score, and model stability. The experimental results are shown in the table below: Table 1. Comparison of the methods of this invention and traditional methods in the classification and identification of ornamental plants.

[0030] As shown in Table 1, the method of this invention significantly outperforms traditional methods in all metrics. The classification accuracy reaches 90.2%, an improvement of approximately 11.8 percentage points compared to CNN and approximately 8.5 percentage points compared to GCN. Precision and recall are improved to 89.5% and 88.7% respectively, indicating that this method can more accurately distinguish categories under complex backgrounds and diverse plant morphologies. The F1 score reaches 89.0%, balancing precision and recall, further demonstrating the robustness of the method. The stability index is 87.6%, still nearly 8 percentage points higher than GCN.

[0031] These performance improvements stem from the invention's differentiated processing of gradient, curl, and harmonic components in the direction-sensitive Hodge decomposition, effectively capturing local variations, loop rotations, and global consistency. Furthermore, the integration of the GSG scattering pyramid enables multi-level recursive expansion, ensuring a unified representation of local and global features. Therefore, even under complex organ morphologies and varying lighting conditions, the classification module maintains high accuracy and stability.

[0032] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for classifying and recognizing ornamental plants based on graph neural networks, characterized in that, Includes the following steps: Images of ornamental plant samples are acquired, image preprocessing is performed, organ regions are segmented by combining morphological features, organ-level graphs are constructed, and node features and edge attributes are generated in the graphs. Based on organ-level graphs, the GSG network is invoked to construct a curve scattering module, which performs multi-scale directional decomposition on node features, generates directional feature representations, and simultaneously generates directional indices. The directional feature representation is mapped to the edge space, and the topological representation is constructed by combining the edge attributes. The connection relationship and directional transmission relationship between nodes are preserved in the topological representation. Perform Hodge decomposition on the topological representation to divide the edge space features in the topological representation into gradient flow branch, curl flow branch and harmonic flow branch; In the gradient flow branch, the curl flow branch, and the harmonic flow branch, GSG scattering pyramids are constructed respectively, and the decomposition depth and scale are expanded layer by layer to extract multi-level flow feature representations; Based on directional feature representation and multi-level flow feature representation, feature alignment and weighted fusion are performed to generate plant-level feature vectors; The plant-level feature vector is input into the classification module, the feature classification operation is performed, the classification probability and category label are output, and the organ region contribution heatmap is generated by combining the directional index.

2. The method for classifying and identifying ornamental plants based on graph neural networks according to claim 1, characterized in that, The generation of the organ-level graph, node features, and edge attributes includes: Based on images of ornamental plant samples, image preprocessing operations are performed, including image denoising, geometric distortion correction, and resolution unification. Based on the preprocessed image, an organ region segmentation method is used to generate organ region segmentation results, extract the node set and edge set, and based on the node set and edge set, an organ-level graph is generated using a graph structure construction method. Based on the organ region segmentation results, morphological features are extracted, a morphological feature set is generated, and the morphological feature set is written into the node features. Based on organ-level graphs and node features, a geometric feature encoding method is used to map the distance, angular relationship, and curvature feature combination between organ nodes into edge attributes.

3. The method for classifying and identifying ornamental plants based on graph neural networks according to claim 1, characterized in that, The generation of the directional feature representation and directional index includes: Based on organ-level maps and node features, the GSG network is invoked to construct a curve scattering module and set multi-scale directional decomposition parameters to generate a set of decomposition parameters. Based on the curve wave scattering module and decomposition parameter set, the curve wave multi-scale directional decomposition method is used to decompose the node features, extract local structural features at different scales and directions, and generate directional feature representations. Based on directional feature representation, the scale parameters and directional parameters generated during the multi-scale directional decomposition of curve waves are written into the index structure to generate a directional index.

4. The method for classifying and identifying ornamental plants based on graph neural networks according to claim 1, characterized in that, The construction of the topological representation includes: Based on the directional feature representation and the edge set of the organ-level graph, the node features are mapped to the edge space to generate the initial edge space features; Based on the initial edge space features and edge attributes, the features are fused to generate fused edge space features; Based on the integration of edge space features, the connection relationships between nodes are constructed; Encode the direction index, pass the direction information to the edge space, and generate the direction transfer relationship; A topological representation is constructed based on the fusion of edge space features, node connection relationships, and direction propagation relationships.

5. The method for classifying and identifying ornamental plants based on graph neural networks according to claim 1, characterized in that, The generation of the three types of structured edge features includes: Based on topological and directional feature representations, edge space features are extracted and discrete exterior differential operators and edge Laplacian operators are constructed to form a set of decomposition operators; Based on the direction index, under the action of the decomposition operator set, the direction-sensitive Hodge decomposition is performed on the edge space features, and the edge space features are decomposed into gradient components, curl components and harmonic components under different directions respectively. Under the direction index, the edge space propagation mechanism is invoked to accumulate the gradient components in the edge set of the topological representation with direction weights. The local gradient values ​​are propagated and smoothed through the direction-related edge weights to generate gradient flow branches. The loop aggregation mechanism is invoked to perform directed cyclic accumulation of curl components in the loop set of topological representation. The rotation propagation intensity is calculated through the edge connection relationship within the loop, and feature transfer paths are established in the direction-related loops to generate curl flow branches. The global consistency mechanism is invoked to project the harmonic components onto the edge Laplacian operator null space in the topological representation. The global projection maintains the directional consistency across organ regions, and a normalization mapping is performed on the direction-dependent overall topological structure to generate harmonic flow branches.

6. The method for classifying and recognizing ornamental plants based on graph neural networks according to claim 1, characterized in that, The generation of the multi-level flow feature representation includes: Based on gradient flow branch, curl flow branch and harmonic flow branch, the number of levels, scale range and number of directional channels of GSG scattering pyramid are set respectively, and scale transfer parameters between levels are established. The gradient component, curl component, and harmonic component are input into the GSG scattering pyramid, and the GSG scattering operation unit is called. In the first layer, the curvelet scattering transformation is performed to extract component features, and the first layer features are output by combining nonlinear activation and aggregation mechanisms. In the higher layers, the lower-level features are recursively input, and the scale transfer parameter is called to control the scale expansion relationship. The curvelet scattering transformation, nonlinear activation, and aggregation are continued until all layers are completed, and the multi-scale hierarchical features of the gradient flow branch, curl flow branch, and harmonic flow branch are obtained respectively. Based on multi-scale hierarchical features of gradient flow branch, curl flow branch and harmonic flow branch, channel alignment, scale alignment and direction index alignment are performed, weighted fusion unit is called, features are weighted and converged at the same level, and layer-by-layer merging is performed between levels to generate multi-level flow feature representations covering local and global aspects.

7. The method for classifying and identifying ornamental plants based on graph neural networks according to claim 1, characterized in that, The generation of the plant-level feature vector includes: Based on directional feature representation and multi-level flow feature representation, channel alignment, scale alignment and directional index alignment are performed to generate an aligned feature set; Based on the aligned feature set, the branch weights and scale weights are calculated using the orientation index to generate a fusion weight set; Based on the aligned feature set and the fusion weight set, different branches are weighted and superimposed at the same level, and layer-by-layer convergence is performed between levels to generate fused features; Based on the fusion features, global aggregation and vectorization mapping are performed to generate plant-level feature vectors.

8. The method for classifying and recognizing ornamental plants based on graph neural networks according to claim 1, characterized in that, The generation of the classification probability, category label, and organ region contribution heatmap includes: Based on plant-level feature vectors, the classification module performs forward inference to generate classification probabilities and determine class labels, while backpropagation is performed to generate gradient information for the classification module. Based on the gradient information of the classification module, the gradient attribution method is used to calculate the contribution of plant-level feature vectors to the directional channel, and the directional contribution weight is generated by combining the directional index. Based on the directional contribution weight and the organ-level graph, the directional contribution weight is propagated along the organ node set and edge set, and then back-projected to the organ region according to the organ node to generate an organ region contribution heatmap.

Images