A similarity measurement method and system for complex landscape

Abstract

The application belongs to the field of tourism recommendation, and particularly relates to a similarity measurement method and system for complex landscapes, which comprises the following steps: obtaining multi-modal data time series of garden landscapes from a landscape training data set and converting the multi-modal data time series into time series vectors; training a feature extraction model according to a multi-modal time series data matrix to obtain an optimal feature extraction model; obtaining an optimal global vector by using the optimal feature extraction model and the multi-modal time series data matrix, and performing decoupling mapping to obtain an attribute vector; inputting the attribute vector and the training data set into a similarity model to optimize the similarity model and obtain an optimal similarity model; inputting actual landscape data into the optimal feature extraction model to obtain an actual global vector; and inputting the actual global vector into the optimal similarity model to obtain the similarity of different landscapes. The application can obtain more accurate similarity between gardens, provide reasonable basis for tourists' travel route planning, and thus better plan the travel route.

Description

Technical Field

[0001] This invention belongs to the field of tourism recommendation, specifically relating to a similarity measurement method and system for complex landscapes. Background Technology

[0002] As an outstanding representative of the Chinese classical garden system, the classical gardens of Jiangnan embody the essence of traditional Chinese philosophy, literature, calligraphy, painting, and architecture, and are renowned for their garden-making philosophy of "though made by man, it appears as if created by nature." Through the organic combination of water features, rocks, architecture, and plants, they create a unique aesthetic effect of ever-changing scenery within a small space. While Jiangnan gardens share a common overall style, different gardens exhibit both obvious similarities in layout, element arrangement, spatial sequence, and cultural themes, yet each possesses its own unique characteristics. Visitors often cannot distinguish which similar gardens they have already visited, leading to visits to highly similar gardens and a lack of immersive experience.

[0003] Recommendations for garden visits mainly rely on human experience or simple labels (such as garden area, main landscape elements, etc.), lacking systematic quantitative analysis of multi-dimensional characteristics such as garden spatial structure, dynamic changes in tour paths, and visual landscape sequences. As a result, it is impossible to accurately let tourists understand the similarity between gardens. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a similarity measurement method and system for complex landscapes, accurately give the similarity between gardens, provide a reasonable basis for tourists to plan their travel routes, and thus better plan their travel routes.

[0005] A similarity measurement method for complex landscapes includes:

[0006] Obtain a landscape training dataset for garden landscapes, and obtain multimodal time series data of garden landscapes based on the landscape training dataset. The training data includes data on changes in scenery, data on the interest of the landscape, data on the distribution of landscape elements, data on changes in the opening and closing of the landscape, and data on the aesthetics of the landscape.

[0007] The multimodal data time series is converted into time series vectors, and multiple time series vectors are combined to form a multimodal time series data matrix;

[0008] Set up a feature extraction model, train the feature extraction model based on the multimodal time series data matrix to obtain the optimal feature extraction model, and obtain the optimal global vector based on the optimal feature extraction model and the multimodal time series data matrix;

[0009] The optimal global vector is decoupled and mapped to obtain the attribute vector;

[0010] The attribute vector and training dataset are input into the similarity model to optimize the similarity model and obtain the optimal similarity model.

[0011] Obtain actual landscape data for different garden landscapes;

[0012] The actual landscape data is input into the optimal feature extraction model to obtain the actual global vectors of different garden landscapes. The actual global vectors of different garden landscapes are then input into the optimal similarity model to obtain the similarity between the different landscapes.

[0013] Optionally, the step of setting the feature extraction model and training the feature extraction model based on the multimodal time-series data matrix to obtain the optimal feature extraction model includes:

[0014] The multimodal time series data matrix is ​​forward encoded and backward encoded to obtain an output vector. A concatenation vector is obtained from the output vector. A normalized weight is obtained from the concatenation vector. The time series features are obtained from the normalized weight and the concatenation vector. The time series features at different time steps form a time series sequence.

[0015] The global vector is obtained based on the time sequence.

[0016] The orthogonal constraint loss is calculated based on the global vector.

[0017] The training data pairs and ground truth similarity are obtained based on the training dataset, and the mean squared error loss is calculated based on the training data pairs and ground truth similarity.

[0018] The reconstruction loss is calculated based on the training data pairs, the multimodal data time series, and the global vector.

[0019] The global vector is decoupled to obtain the attribute vector, and the threshold penalty loss is calculated based on the attribute vector and the training data pair.

[0020] The threshold penalty loss, reconstruction loss, mean squared error loss, and orthogonal constraint loss are weighted and summed to obtain the total loss. The model parameters with the minimum total loss are then obtained to obtain the optimal feature extraction model.

[0021] Optionally, the multimodal time series data matrix is ​​subjected to forward encoding and backward encoding respectively to obtain an output vector, a concatenation vector is obtained based on the output vector, a normalized weight is obtained based on the concatenation vector, and a time series feature is obtained based on the normalized weight and the concatenation vector. The time series features at different time steps form a time series sequence including:

[0022] The multimodal time-series data matrix is ​​encoded using forward LSTM to obtain the forward output vector;

[0023] The multimodal time-series data matrix is ​​then encoded using inverse LSTM to obtain an inverse output vector.

[0024] The forward output vector and the reverse output vector are concatenated to obtain the concatenated vector;

[0025] Calculate the attention score for each concatenated vector and normalize the attention score to obtain the normalized weight;

[0026] The time series features are obtained based on the normalized weights and the concatenated vectors, and the time series features at different time steps form a time series sequence.

[0027] Optionally, obtaining the global vector based on the time series includes:

[0028] The time series is processed by multi-head self-attention to obtain projection features;

[0029] The time series is processed by multi-head self-attention to obtain projection features, and multi-head splicing features are obtained based on the projection features.

[0030] Add residual connections to the multi-head splicing features to obtain residual features, and normalize the residual features to obtain normalized residual features;

[0031] Calculate the attention score of the residual normalized feature, and obtain the feature weight based on the attention score;

[0032] The global vector is obtained based on the feature weights and attention scores.

[0033] Optionally, the step of inputting the attribute vector and the training dataset into the similarity model to optimize the similarity model and obtain the optimal similarity model includes:

[0034] The attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector.

[0035] Mahalanobis distance, mutual information, and attribute entropy are calculated based on the attribute vectors of different landscapes.

[0036] The Euclidean distance is calculated based on the attribute vectors of different landscapes, and the variation similarity is obtained based on the Euclidean distance and the attribute vectors of different landscapes.

[0037] A comprehensive similarity is obtained based on the Mahalanobis distance, attribute entropy, mutual information, and change similarity, and a similarity loss is obtained based on the comprehensive similarity.

[0038] The reconstruction loss is obtained based on the training data.

[0039] Based on the reconstruction loss, similarity loss, and attribute loss, the total similarity loss is obtained. The similarity model is then optimized based on the total similarity loss to obtain an optimized similarity model.

[0040] Optionally, the attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector, and is expressed as:

[0041]

[0042] in, For attribute loss, For the first The attribute threshold corresponding to each attribute vector. For the first A vector of attributes The total number of attributes. For the first The values ​​of the attribute vector.

[0043] Optionally, the comprehensive similarity obtained based on the Mahalanobis distance, attribute entropy, mutual information, and change similarity, and the similarity loss obtained based on the comprehensive similarity, are expressed as follows:

[0044]

[0045] in, , , as well as As weight, The Mahalanobis distance, For the similarity of changes, For mutual information, This is the attribute entropy.

[0046] A similarity measurement system for complex landscapes includes:

[0047] The data acquisition module is used to acquire a landscape training dataset of the garden landscape, and to acquire multimodal data time series of the garden landscape based on the landscape training dataset. The training data includes data on changes in scenery, data on the interest of the landscape, data on the distribution of landscape elements, data on changes in the opening and closing of the landscape, and data on the aesthetics of the landscape.

[0048] The data conversion module is used to convert the multimodal data time series into time series vectors, and multiple time series vectors form a multimodal time series data matrix;

[0049] The feature extraction model training module is used to set up the feature extraction model, train the feature extraction model based on the multimodal time series data matrix to obtain the optimal feature extraction model, and obtain the optimal global vector based on the optimal feature extraction model and the multimodal time series data matrix.

[0050] The decoupling mapping module is used to decouple and map the optimal global vector to obtain the attribute vector;

[0051] The similarity model optimization module is used to input the attribute vector and training dataset into the similarity model to optimize the similarity model and obtain the optimal similarity model.

[0052] The actual data acquisition module is used to acquire actual landscape data for different garden landscapes;

[0053] The similarity calculation module is used to input the actual landscape data into the optimal feature extraction model to obtain the actual global vectors of different garden landscapes, and input the actual global vectors of different garden landscapes into the optimal similarity model to obtain the similarity of different landscapes.

[0054] A terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a similarity measurement method for complex landscapes.

[0055] A computer-readable storage medium storing a computer program that, when loaded and executed by a processor, employs a similarity measurement method for complex landscapes.

[0056] The beneficial effects of this invention are:

[0057] Five temporal aesthetic features were defined: scenic variation, interest, element distribution, opening and closing variations, and aesthetic themes, forming a multimodal input. A high-order dynamic aesthetic representation of the gardens was extracted using a deep network incorporating modality-specific LSTM encoding, cross-modal temporal attention fusion, temporal self-attention, and global pooling. An innovative attribute non-compensation mechanism was introduced to ensure that each aesthetic attribute contributes independently and that weaknesses cannot be compensated for by strengths. This mechanism was used to calculate the similarity of complex landscapes, resulting in more accurate similarity between gardens and providing a reasonable basis for tourists' route planning, thus enabling better tourism route planning. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating a similarity measurement method for complex landscapes according to the present invention.

[0059] Figure 2 This is a schematic diagram of ten representative complex landscape similarity matrices of the present invention;

[0060] Figure 3 This is a schematic diagram of ten representative complex landscape similarity clusters of the present invention;

[0061] Figure 4 This is a schematic diagram of the loss curve during the model training process of the present invention;

[0062] Figure 5 This is a schematic diagram of the complex landscape similarity calculation results of the present invention;

[0063] Figure 6 This is the similarity MDS diagram of the present invention;

[0064] Figure 7 This is a schematic diagram illustrating the dynamic characteristics analysis of the tour route according to the present invention;

[0065] Figure 8 This is a schematic diagram illustrating the similarity analysis between each garden in this invention and known complex landscapes. Detailed Implementation

[0066] A similarity measurement method for complex landscapes, such as Figure 1 As shown, it includes:

[0067] S1. Obtain the landscape training dataset of the garden landscape. Based on the landscape training dataset, obtain the multimodal data time series of the garden landscape. The training data includes data on changes in scenery, data on the interest of the landscape, data on the distribution of landscape elements, data on changes in the opening and closing of the landscape, and data on the aesthetics of the landscape.

[0068] Among them, the multimodal data time series includes the sequence of unusual scenery, the sequence of interesting content, the sequence of element distribution, the sequence of opening and closing changes, and the sequence of aesthetic themes.

[0069] Specifically, the sequence of different scenic views is mainly based on the distribution changes of multiple categories of visual elements in continuous scene images along the tour route. The original data is generally the recognition results of multiple categories of visual elements in the landscape map, such as buildings, water bodies, plants, rocks, etc.

[0070] Interest sequences are derived from local visual information density, rhythm changes, and global baseline statistics, and are used to characterize the attractiveness of the tour experience. The original data generally comes from tourist questionnaire statistics, which is the rhythm intensity sequence of the tourist's perceived experience.

[0071] The element distribution sequence relies on the garden plan and spatial annotation data to reflect the spatial organization changes of elements such as buildings, plants, and water bodies along the path. The original data comes from the measured garden plan and the spatial annotation data of buildings, plants, water bodies, rocks, etc.

[0072] The opening and closing change sequence perceives the degree of spatial enclosure and openness from scene images. The raw data comes from scene images along the tour path and the degree of angular change of the route.

[0073] The aesthetic theme sequence is extracted from the landscape text description by natural language processing methods and mapped into temporal features. The original data is landscape text description, including historical and cultural interpretations.

[0074] Specifically, multimodal data during the landscape tour is acquired through sensors, image acquisition, audio recording, and text description. The tour path is discretized into multiple time points according to the time step. For each time point, five modal feature vectors are extracted and aggregated into a scalar sequence through feature engineering. The multimodal temporal network aims to process temporal data on complex landscape element distribution, opening and closing variations, and aesthetic themes, extracting high-level features through deep learning. The optimized network architecture includes a multimodal data fusion layer, a temporal feature extraction layer, and an attribute decoupling representation layer. The multimodal data fusion layer, considering modeling metrics, designs the input data for the multimodal temporal network to include five modal temporal sequences: sequences of varied landscape changes, sequences of interesting features, sequences of element distribution, sequences of opening and closing variations, and sequences of aesthetic themes. Detailed modeling is as follows:

[0075] Sequence of Amazing Scenes:

[0076] in, For the rate of change of landscape type, This indicates the landscape type label corresponding to the visitor's location at time t. The landscape type is determined by a predefined set of landscape element categories. The information entropy represents the distribution of landscape elements at time t, where c is the index of the landscape element category. This represents the probability of the c-th type of landscape element appearing at time t, used to characterize the degree of diversity in the local landscape composition. For the surprise of the transition matrix, The length distribution of the stable segment is given by t, where t is time.

[0077] Interesting sequence: in, For local interest density, As a baseline for global interest, The relative rate of change This is the rhythm smoothing factor.

[0078] Element distribution sequence: ,in, For spatial entropy, For permeability, For the scale of plant aggregation and dispersion, For shape diversity.

[0079] Opening and closing change sequence: ,in, The average opening and closing index, As a rhythm smoothing factor, To ensure suitability, This represents the maximum rate of change.

[0080] Aesthetic theme sequence: Extracting topic tags through natural language processing. This refers to the aesthetic theme tags associated with the visitor location or landscape node corresponding to time t. The aesthetic themes are obtained by natural language processing of landscape text descriptions, historical and cultural introductions, or expert-annotated texts. The set of theme tags is determined by a pre-constructed aesthetic theme thesaurus. () is a one-hot encoding function used to map discrete aesthetic topic labels into binary vectors of fixed dimensions. The dimension is equal to the total number of aesthetic topic categories. When the dimension corresponding to the current topic is 1, the other dimensions are 0.

[0081] Each modality is encoded using an independent bidirectional LSTM to extract temporal dynamic features. The hidden states of the forward and backward LSTMs are concatenated to form a complete representation for each time step. The complex landscape modality data sequences described above are then unified, and we set the... The tour routes of the gardens are discretized into Each time step, each time step The data includes observations from five modalities. The original observation for each modality is a multidimensional feature vector, which is aggregated into a scalar through feature engineering and then processed into a vector representation.

[0082]

[0083] in, For the first A garden, For the first Each mode in time The original feature vector (i.e., obtained through sequence processing). The values ​​from 1 to 5 correspond to the following respectively. It is an aggregate function. These are modal indexes, corresponding to different scenes, interesting elements, element distribution, opening and closing variations, and aesthetic themes.

[0084] S2. Transform the multimodal time series data into time series vectors, and combine multiple time series vectors to form a multimodal time series data matrix;

[0085] Specifically, the complete multimodal time series data of each garden can be represented as a matrix. This represents a complete multimodal time series data matrix with dimension T. 5, specifically expressed as:

[0086]

[0087] in, , Let T be the set of real numbers, and T be the transpose of the set.

[0088] S3. Set up a feature extraction model, train the feature extraction model based on the multimodal time series data matrix to obtain the optimal feature extraction model, and obtain the optimal global vector based on the optimal feature extraction model and the multimodal time series data matrix.

[0089] A feature extraction model is set up and trained using a multimodal time-series data matrix to obtain the optimal feature extraction model, including:

[0090] The multimodal time series data matrix is ​​forward encoded and backward encoded to obtain output vectors. A concatenated vector is obtained from the output vectors. Normalized weights are obtained from the concatenated vectors. Temporal features are obtained from the normalized weights and the concatenated vectors. Temporal features at different time steps form a time series.

[0091] The multimodal time series data matrix is ​​forward encoded and backward encoded to obtain output vectors. A concatenated vector is obtained from the output vectors. Normalized weights are obtained from the concatenated vectors. Temporal features are obtained from the normalized weights and the concatenated vectors. The temporal features at different time steps form a time series sequence.

[0092] The multimodal time-series data matrix is ​​encoded using forward LSTM to obtain the forward output vector;

[0093] The multimodal time-series data matrix is ​​encoded using inverse LSTM to obtain the inverse output vector;

[0094] Specifically, each mode is encoded using an independent bidirectional LSTM to capture mode-specific temporal patterns. The forward LSTM update formula is defined as follows:

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101] in, Forward input gate vector, Forward forget gate vector, This is the forward output gate vector. , , , Forward weight vector, , , as well as This is the forward bias vector. For activation function, This is a hyperbolic tangent activation function. This represents element-wise multiplication, and f represents the forward LSTM encoding process. This represents the forward candidate memory cell state. This is the forward memory cell state. In the forward-hidden state, for The forward hidden state at any given moment.

[0102] Define a backward LSTM to process the sequence in reverse, as follows:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] in, The inverse input gate vector, This is the reverse forgetting gate vector. This is the inverse output gate vector. 、 、 as well as The reverse weight vector, 、 、 as well as It is the reverse bias vector. This represents the reverse candidate memory cell state. This is a reverse memory cell state. In reverse hidden state, for The reverse hidden state at any given moment.

[0110] The forward output vector and the backward output vector are concatenated to obtain the concatenated vector;

[0111] Specifically, the splicing is represented as:

[0112]

[0113] in, In the forward-hidden state, In reverse hidden state, This is a concatenated vector.

[0114] Calculate the attention score for each concatenated vector and normalize the attention score to obtain the normalized weight;

[0115] The temporal features are obtained by normalizing the weights and concatenating the vectors, and the temporal features at different time steps form a temporal sequence.

[0116] The global vector is obtained from the time sequence;

[0117] Specifically, information from five modalities is dynamically fused through an attention mechanism, and the attention weight of each modality is calculated. The fused features are further processed by multi-head self-attention to extract temporal dependencies, and finally, the global feature vector is extracted through temporal attention pooling.

[0118] Specifically, the cross-modal attention mechanism aims to dynamically fuse information from five modalities. First, it calculates the time signature of each modality. Attention score:

[0119]

[0120] in, The attention score for the concatenated vectors. 、 、 and For learnable parameters, For cross-modal averaging, ReLU is the activation function, and T is the transpose.

[0121] Normalizing the attention scores yields the normalized weights, expressed as:

[0122]

[0123] in, For normalized weights, Indicates the time t at the first hour The unnormalized attention score corresponds to each modality. The attention score is calculated by combining the temporal feature representation of that modality with the cross-modal average representation. The modal index variable corresponds to five modalities: scene variation, interest, element distribution, opening and closing variation, and aesthetic theme.

[0124] The temporal features obtained from the normalized weights and concatenated vectors are expressed as follows:

[0125]

[0126] in, For time-series features, N is the number of modes. A mode-specific projection matrix.

[0127] The global vector is obtained from the time series, including:

[0128] The time series is processed by multi-head self-attention to obtain projection features, and multi-head splicing features are obtained based on the projection features.

[0129] Add residual connections to the multi-head splicing features to obtain residual features, and normalize the residual features to obtain normalized residual features;

[0130] Calculate the attention score of the residual normalized feature, and obtain the feature weight based on the attention score;

[0131] The global vector is obtained based on the feature weights and attention scores.

[0132] Specifically, multi-head self-attention pairs fusion sequences Processing is required for the first... The size is then subjected to linear projection processing, as follows:

[0133]

[0134]

[0135]

[0136] in, 、 and For query matrix (projection features). 、 and for The projection matrix of the h-th head.

[0137] Define scaled dot product attention as follows:

[0138]

[0139] in, The dimension of the sequence. This is for scaling dot product attention.

[0140] The query matrix of the h-th attention head Key matrix Sum matrix The attention output of the h-th head is calculated and expressed as:

[0141]

[0142]

[0143] in, The output projection matrix is ​​used to linearly map the concatenated feature representations of each attention point back to the unified feature space dimension of the model. The output projection matrix is ​​a learnable parameter, obtained during model training through backpropagation and optimization along with other network parameters. The input is the temporal feature matrix of the multi-head self-attention module.

[0144] After MHSA, we add residual connections and layer normalization, represented as follows:

[0145]

[0146] in, This is a residual normalization feature.

[0147] To extract the global representation from the sequence, we use temporal attention pooling. First, we calculate the attention score for each time step, expressed as:

[0148]

[0149] in, , and For learning parameters, The attention score is the feature of the residual normalization.

[0150] The feature weights are obtained from the attention scores and are expressed as follows:

[0151]

[0152] in, This is a time step index variable used to iterate through the entire time series. Indicates time step The corresponding unnormalized temporal attention score measures the relative importance of the feature at that time step in the global representation construction process. The attention score is calculated by using a learnable attention scoring function to represent the temporal feature at that time step. The feature weight represents the normalized feature weight corresponding to time step t. It is obtained by performing Softmax normalization on the attention scores of all time steps, so that the sum of the weights of each time step is 1, thereby realizing adaptive weighting of features at different times.

[0153] Based on the feature weights and attention scores, the global vector is obtained, represented as:

[0154]

[0155] in, For the first A time series, It is a global vector.

[0156] The orthogonal constraint loss is calculated based on the global vector.

[0157] Specifically, the global vector Mapping to K independent aesthetic attribute spaces, orthogonal constraints ensure the independence between attributes. Each attribute corresponds to an aesthetic dimension. Further modeling of attribute decoupling projection is performed; in this embodiment, let... Each aesthetic attribute is represented by a decoupled projection matrix P, which maps the global representation to the attribute space. Let P be the attribute vector. P is the decoupling projection matrix, expressed as:

[0158]

[0159] in, For different aesthetic attribute vectors.

[0160] The orthogonal constraint loss is expressed as:

[0161]

[0162] in, For orthogonal constraint loss, To decouple the projection matrix of the first List, To decouple the projection matrix of the first List, It is the Kronecker delta function.

[0163] The training data pairs and ground truth similarity are obtained from the training dataset, and the mean squared error loss is calculated based on the training data pairs and ground truth similarity.

[0164] The reconstruction loss is calculated based on the training data pairs, the multimodal data time series, and the global vector.

[0165] The global vector is decoupled to obtain the attribute vector. The threshold penalty loss is calculated based on the attribute vector and the training data pair.

[0166] The threshold penalty loss, reconstruction loss, mean squared error loss, and orthogonal constraint loss are weighted and summed to obtain the total loss. The model parameters that minimize the total loss are then obtained to obtain the optimal feature extraction model.

[0167] Specifically, the mean squared error loss is expressed as:

[0168]

[0169] in, For gardens He Garden The ground's true similarity, For gardens With gardens The overall similarity between them For the training data pair set. The size of the training data pairs.

[0170] The reconstruction loss is expressed as:

[0171]

[0172] in, For norm, Let be the global vector of the k-th garden. Let represent the original multimodal time-series input feature matrix corresponding to the k-th garden, which is composed of the multimodal time series obtained by discretizing the garden along the tour path. Its specific form is as follows:

[0173] =

[0174] in, 2 represents the number of time steps, and M represents the number of modes (M=5 in this embodiment). For the first Each garden in time t The scalar observation value corresponding to the m-th mode. As the reconstruction target of the decoder, it is used to constrain the global representation. Preserve the original multimodal timing information. Let be the set of real numbers. The threshold penalty loss is expressed as:

[0175]

[0176] in, For the k-th original attribute vector, The minimum threshold is set for each attribute. When the attribute value is below this threshold, it will contribute to the loss function. The 90th percentile is usually chosen to define the threshold.

[0177] The total loss is expressed as:

[0178]

[0179] in, This is the set of model parameters for the feature extraction model. , and The weights are used to determine the optimal model parameters by minimizing the total loss, thus obtaining the optimal feature extraction model.

[0180] S4. Decouple and map the optimal global vector to obtain the attribute vector;

[0181] S5. Input the attribute vector and training dataset into the similarity model to optimize the similarity model and obtain the optimal similarity model;

[0182] The attribute vectors and training dataset are input into the similarity model to optimize it, resulting in the optimal similarity model, which includes:

[0183] The attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector.

[0184] The attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector, and is expressed as:

[0185]

[0186] in, For attribute loss, For the first The attribute threshold corresponding to each attribute vector. For the first A vector of attributes The total number of attributes. For the first The values ​​of the attribute vector.

[0187] Attribute values ​​are calculated based on the attribute vectors of different landscapes, and non-compensatory similarity is obtained based on the attribute values ​​of different landscapes, including:

[0188] Attribute values ​​are calculated based on the attribute vectors of different landscapes;

[0189] Cosine similarity is calculated based on the attribute vectors of different landscapes;

[0190] The non-compensatory similarity is calculated based on the cosine similarity and attribute values.

[0191] Mahalanobis distance, mutual information, and attribute entropy are calculated based on the attribute vectors of different landscapes.

[0192] The Euclidean distance is calculated based on the attribute vectors of different landscapes, and the variation similarity is obtained based on the Euclidean distance and the attribute vectors of different landscapes.

[0193] The comprehensive similarity is obtained based on Mahalanobis distance, attribute entropy, mutual information, and non-compensatory similarity, and the similarity loss is obtained based on the comprehensive similarity.

[0194] The reconstruction loss is obtained based on the training data;

[0195] Based on the reconstruction loss, similarity loss, and attribute loss, the total similarity loss is obtained. The similarity model is then optimized based on the total similarity loss to obtain the optimized similarity model.

[0196] Specifically, the non-compensatory attribute mechanism states that in a cultural landscape system, a weakness or deficiency in a particular attribute cannot be compensated for by the system's strengths in other attributes. This mechanism manifests as the relative independence of the numerical values ​​of different attributes. Each attribute plays an indispensable role in achieving the overall function and cannot be simply replaced or compensated for by the performance of other attributes. The relationships between attributes cannot be viewed as linear or additive, but rather as a complex interdependence and relative independence. When designing and evaluating a cultural landscape system, special attention needs to be paid to the comprehensiveness and balance of each attribute to ensure that no attribute weakness affects the overall functional realization.

[0197] Attribute values ​​are calculated based on the attribute vectors of different landscapes, and non-compensatory similarity is obtained based on the attribute values ​​of different landscapes.

[0198] Traditional distance-based methods can assess the overall characteristics of complex classical landscapes in Jiangnan and select the scheme closest to the ideal solution. Data statistics methods based on correlation measures can assess the similarity of feature differences across different attributes and select complex landscapes with trends similar to the ideal solution. Information entropy-based measures quantify the information content and uncertainty of different attributes in complex landscapes, providing a reference for assessing the overall similarity of complex landscapes. High entropy values ​​can represent the diversity and complexity of attribute performance, reflecting the attribute capabilities of complex landscapes under different environments. Correlation-based measures represent the strength and direction of the correlation between two or more attributes, measuring their interrelationship. The attribute non-compensatory mechanism emphasizes that the weakness of any attribute cannot be compensated for by the strength of other attributes. This model achieves this through loss function constraints and a threshold mechanism in similarity calculation.

[0199] A minimum threshold is set for each attribute; if an attribute fails to meet the threshold, the overall score is significantly reduced in similarity calculation. A threshold penalty term is introduced into the loss function to enhance sensitivity to weaker attributes during training. In similarity calculation, a weaker attribute weight is introduced to emphasize the weakest link effect.

[0200] Define a threshold function for multimodal attributes in complex landscapes to ensure that each attribute meets the minimum requirement, expressed as:

[0201]

[0202] in, For the first Thresholds for attribute vectors.

[0203] Define an irreplaceable similarity calculation method, embed an irreplaceable mechanism into the similarity calculation, and calculate the irreplaceable similarity, expressed as:

[0204]

[0205]

[0206] in, For Garden A's i A vector of attributes Let be the i-th attribute vector of garden B. For cosine similarity, and As weight, This represents the value of the attribute vector. The first item... This ensures that only complex landscapes meeting all attribute criteria achieve high similarity, thus achieving non-compensatory similarity. The second point emphasizes the decisive role of the weakest attribute, with its weighting... Calculated using the entropy weight method To adjust the parameters.

[0207] By combining multimodal data, this study characterizes the similarity of complex landscapes, analyzes their shared aesthetic features and universal patterns, and constructs a multi-dimensional similarity measurement system. This system integrates distance, similarity, association, and entropy measures to comprehensively assess the similarity of complex landscapes. The distance measure—overall proximity—uses Mahalanobis distance, considering the covariance between attributes. The attribute covariance matrix is ​​estimated using training data and calculated as follows:

[0208] Mahalanobis distance is calculated as follows:

[0209]

[0210] in, This is the Mahalanobis distance.

[0211] The similarity of changes is calculated as follows:

[0212]

[0213]

[0214] in, For Euclidean distance, For variation similarity.

[0215] Mutual information is used to quantify the strength of the association between attributes. High mutual information indicates strong synergy between attributes, expressed as:

[0216]

[0217] in, Let be the i-th attribute vector of garden A. Let j be the attribute vector of garden B. The first garden A Marginal probability distribution of attribute vectors in attribute space Indicates the first of garden B The marginal probability distribution of an attribute vector in the attribute space is used to characterize the statistical characteristics of the attribute value in the overall sample. For Garden A's The attribute and the first garden B The joint probability distribution of two attributes is used to characterize the joint occurrence relationship of two attributes in the sample space.

[0218] The attribute entropy is calculated as follows:

[0219]

[0220] Where s is the time scale. For scale weights, For attribute entropy, This is the index of the value state of this attribute after discretization. For attribute index, indicating the first One decoupling attribute, Indicates the first The first attribute A state vector with a value.

[0221] The overall similarity is calculated as follows:

[0222]

[0223] in, , , as well as As the weight, in this embodiment, =0.4, =0.3, =0.2, =0.1.

[0224] To accelerate the solution process, this paper details the training strategy and optimization. For the loss function design, a multi-task learning framework is employed, integrating various constraints. ,in To reconstruct the loss, the constraint encoding-decoding process retains as much structural information of the original multimodal time series data as possible while compressing the representation, thereby ensuring the integrity of the global vector's representation of the input features. X represents the true multimodal time series feature matrix of the model input, which is the original input data of the encoder, consisting of multimodal time series obtained by discretizing the garden along the tour path. This represents the multimodal temporal feature matrix reconstructed by the decoder from the global vectors, and its calculation method is as follows: , where g is the global representation vector obtained through multimodal temporal encoding, attention fusion, and global pooling. This ensures consistency between encoding and decoding. The similarity loss is intended to promote the convergence of similar complex landscape features.

[0225] S6. Obtain actual landscape data for different garden landscapes;

[0226] S7. Input the actual landscape data into the optimal feature extraction model to obtain the actual global vectors of different garden landscapes. Input the actual global vectors of different garden landscapes into the optimal similarity model to obtain the similarity of different landscapes. Specific implementation examples:

[0228] Five types of multimodal time-series data were generated for each complex landscape along the path. Each type of data contains a 10-dimensional feature vector with 100 time points, and the resulting data is shown in Table 1.

[0229] Table 1. Statistical analysis of temporal data characteristics of various complex landscapes

[0230]

[0231] By combining multimodal data from ten representative complex landscapes, this study characterizes the similarity of classical complex landscapes in the Jiangnan region, analyzes their shared aesthetic features and universal patterns, such as... Figure 2 The figure shows ten similarity matrices representing complex landscapes. The non-self-similarity among complex landscapes ranges from 0.7097 to 0.8142, with no cases showing significantly low or extremely high similarity. In the experiment, the four highly similar complex landscape combinations all exhibited regional proximity or stylistic origins: Humble Administrator's Garden-Qiyuan Garden, Yiyuan Garden-Zhanyuan Garden, Qiuxia Garden-Qiyuan Garden, and Couple's Garden-Shenyuan Garden. The Suzhou complex landscape cluster—Humble Administrator's Garden, Lingering Garden, Couple's Garden, and Yiyuan Garden—had an internal mean similarity of 0.772, significantly higher than the mean of the external complex landscapes. Figure 3 For complex landscape similarity clustering, the four types of indicators with high similarity combinations all show synergistic optimization characteristics, with no single indicator dominating, proving that the weight setting of 4:3:2:1 can balance the distribution differences and temporal trends.

[0232] For the similarity model in this paper, the verification complex landscape 1 is set as Tuisi Garden.

[0233] This model is trained using the AdamW optimizer, with an initial learning rate set to... Furthermore, a cosine annealing strategy is used to dynamically adjust the learning rate to improve convergence stability. For regularization, gradient clipping is used to prevent gradient explosion, and the weight decay coefficient is... To control model complexity, Dropout was introduced into the LSTM and attention layers to enhance generalization ability. Training data was derived from ten representative complex landscape scenes, covering multimodal temporal information, and divided into training, validation, and test sets in a 7:2:1 ratio. The batch size was set to 32, and the entire training process consisted of 100 rounds to ensure the model could fully learn and generalize. Figure 4 As shown.

[0234] From the experimental results Figure 5 and Figure 6 It can be seen that the average similarity of Tuisi Garden with the complex landscapes of Zhejiang and Shanghai (0.793) is higher than that with the complex landscapes of Suzhou (0.765). In the MDS diagram, it is classified into the complex landscape cluster of Zhejiang and Shanghai (within the orange ellipse), consistent with the style blending caused by the actual geographical location. The similarity between the newly added complex landscape Tuisi Garden and the existing complex landscapes (0.7215~0.8015) highly overlaps with the similarity range of the existing complex landscapes, indicating that the model did not overfit the original data and the output for the new data conforms to the expected distribution. The contribution ratios of the four indicators for the similarity of the newly added complex landscape Tuisi Garden are consistent with those of the existing complex landscapes, with no single indicator dominating, demonstrating the stability of the calculation logic of the complex landscape similarity model designed in this paper.

[0235] For the path model in this paper, we set the verification complex landscape 2 as Ge Garden in Yangzhou.

[0236] This paper uses a morphological skeleton extraction algorithm to identify the main tour routes within the park and abstracts them into a hypernetwork containing nodes. and corresponding edge sets The topology map. Based on the reconstructed topology network and element layout, we generated a multi-dimensional feature sequence depicting the dynamic experience of the visitor along the path using a sliding window with a sampling step size of 5 meters. This sequence consists of two core components that jointly characterize the visitor's spatiotemporal experience during the journey:

[0237]

[0238] Among them, the degree of wonder Entropy calculations are performed on the probabilities of multiple visual elements to quantify the diversity and variability of the landscape; opening and closing degree. The perception of spatial enclosure and openness is quantified. A deep learning model is used to perceive the opening and closing attributes of spatial forms, obtained through feature extraction from scene images using a pre-trained ResNet model. Figure 7 It can be clearly observed that the dramatic fluctuations in the opening and closing sequence reflect the rich rhythm of the transformation of complex landscape spaces from corridors to courtyards.

[0239] We input this sequence into a trained similarity model and conducted rigorous validation at three levels. First, in the similarity clustering analysis, we calculated the weighted Mahalanobis distance between Geyuan and ten known complex landscapes in the database. For example... Figure 8 The results show that Geyuan Garden and Jichang Garden have the highest similarity (D=0.80), both exhibiting high variance and similar autocorrelation structures in their sequences. This accurately captures their shared design philosophy, which centers on ingenious rockery and path organization and pursues ever-changing scenery with each step. Conversely, Geyuan Garden is the most dissimilar to the Humble Administrator's Garden, which is dominated by open water features (D=1.23), and their characteristic sequences show fundamental differences.

[0240] To conduct a final quantitative evaluation of this generalization validation, we comprehensively evaluated the accuracy of the data reconstruction. Consistency in similarity calculation Path planning rationality The three indicators construct a comprehensive generalization index. Calculations showed that the G-index reached 0.87, significantly higher than the effective generalization threshold of 0.7. Simultaneously, the KL divergence of the model output probability distribution based on the reconstructed and measured data was only 0.043, further demonstrating from an information theory perspective that the data reconstruction process is highly fidelity. Through independent validation on a specific garden, the proposed complex landscape similarity model exhibits strong generalization ability in unstructured data input scenarios, effectively capturing the essential spatial characteristics of complex landscapes and generating reasonable similarity calculation results.

[0241] A similarity measurement system for complex landscapes includes:

[0242] The data acquisition module is used to acquire the landscape training dataset of the garden landscape. Based on the landscape training dataset, it acquires the multimodal data time series of the garden landscape. The training data includes data on changes in scenery, data on the interest of the landscape, data on the distribution of landscape elements, data on changes in the opening and closing of the landscape, and data on the aesthetics of the landscape.

[0243] The data transformation module is used to convert multimodal time series data into time series vectors, and multiple time series vectors form a multimodal time series data matrix.

[0244] The feature extraction model training module is used to set up the feature extraction model, train the feature extraction model based on the multimodal time series data matrix to obtain the optimal feature extraction model, and obtain the optimal global vector based on the optimal feature extraction model and the multimodal time series data matrix.

[0245] The decoupling mapping module is used to decouple and map the optimal global vector to obtain the attribute vector;

[0246] The similarity model optimization module is used to input attribute vectors and training datasets into the similarity model to optimize the similarity model and obtain the optimal similarity model.

[0247] The actual data acquisition module is used to acquire actual landscape data for different garden landscapes;

[0248] The similarity calculation module is used to input actual landscape data into the optimal feature extraction model to obtain the actual global vectors of different garden landscapes, and input the actual global vectors of different garden landscapes into the optimal similarity model to obtain the similarity between different landscapes.

[0249] This application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a similarity measurement method for complex landscapes.

[0250] The terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server. The terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0251] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0252] The memory can be an internal storage unit of the terminal device, such as a hard disk or RAM of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) equipped on the terminal device. Furthermore, the memory can be a combination of internal storage units and external storage devices of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0253] In this terminal device, a similarity measurement method for complex landscapes described in the above embodiments is stored in the terminal device's memory and loaded and executed on the terminal device's processor for convenient use.

[0254] This application also discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it employs a similarity measurement method for complex landscapes as described in the above embodiments.

[0255] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0256] In this computer-readable storage medium, a similarity measurement method for complex landscapes according to the above embodiments is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0257] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0258] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A similarity measurement method for complex landscapes, characterized in that, include: A landscape training dataset for the garden landscape is obtained. Based on the landscape training dataset, multimodal time series data of the garden landscape are obtained. The training data includes data on changes in scenery, data on landscape interest, data on the distribution of landscape elements, data on changes in the opening and closing of landscape features, and data on landscape aesthetics. The sequence of the degree of scenery variation is used as the data on changes in scenery variation, and the sequence of the degree of scenery variation is represented as follows: ,in, For the rate of change of landscape type, This indicates the landscape type label corresponding to the visitor's location at time t. The landscape type is determined by a predefined set of landscape element categories. The information entropy represents the distribution of landscape elements at time t, where c is the index of the landscape element category. This represents the probability of the c-th type of landscape element appearing at time t, used to characterize the degree of diversity in the local landscape composition. For the surprise of the transition matrix, To establish a stable segment length distribution, t represents time; the aesthetic theme sequence is used as landscape aesthetic data, and the aesthetic theme sequence is represented as follows: Extracting topic tags through natural language processing. This refers to the aesthetic theme tags associated with the visitor's location or scenic node at time t. The aesthetic themes are obtained through natural language processing of landscape text descriptions, historical and cultural introductions, or expert-annotated texts. The set of theme tags is determined by a pre-constructed aesthetic theme thesaurus. () is a one-hot encoding function used to map discrete aesthetic topic labels into binary vectors of fixed dimensions, the dimension of which is equal to the total number of aesthetic topic categories. The dimension corresponding to the current topic is set to 1, and the other dimensions are set to 0. The multimodal data time series is converted into time series vectors, and multiple time series vectors are combined to form a multimodal time series data matrix; A feature extraction model is set up, and the feature extraction model is trained based on the multimodal time series data matrix to obtain the optimal feature extraction model. Based on the optimal feature extraction model and the multimodal time series data matrix, the optimal global vector is obtained. The optimal global vector is decoupled and mapped to obtain the attribute vector; The attribute vector and training dataset are input into the similarity model to optimize the similarity model and obtain the optimal similarity model. Obtain actual landscape data for different garden landscapes; The actual landscape data is input into the optimal feature extraction model to obtain the actual global vectors of different garden landscapes. The actual global vectors of different garden landscapes are then input into the optimal similarity model to obtain the similarity between the different landscapes. The step of inputting the attribute vector and training dataset into the similarity model to optimize the similarity model and obtain the optimal similarity model includes: The attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector. Mahalanobis distance, mutual information, and attribute entropy are calculated based on the attribute vectors of different landscapes. The Euclidean distance is calculated based on the attribute vectors of different landscapes, and the variation similarity is obtained based on the Euclidean distance and the attribute vectors of different landscapes. A comprehensive similarity is obtained based on the Mahalanobis distance, attribute entropy, mutual information, and change similarity, and a similarity loss is obtained based on the comprehensive similarity. Obtain training data pairs based on the training dataset, and obtain reconstruction loss based on the training data pairs; Based on the reconstruction loss, similarity loss, and attribute loss, the total similarity loss is obtained. The similarity model is then optimized based on the total similarity loss to obtain an optimized similarity model.

2. The similarity measurement method for complex landscapes as described in claim 1, characterized in that, The step of setting up a feature extraction model, training the feature extraction model based on the multimodal time-series data matrix to obtain the optimal feature extraction model, includes: The multimodal time series data matrix is ​​forward encoded and backward encoded to obtain an output vector. A concatenation vector is obtained from the output vector. A normalized weight is obtained from the concatenation vector. The time series features are obtained from the normalized weight and the concatenation vector. The time series features at different time steps form a time series sequence. The global vector is obtained based on the time sequence. The orthogonal constraint loss is calculated based on the global vector. The training data pairs and ground truth similarity are obtained based on the training dataset, and the mean squared error loss is calculated based on the training data pairs and ground truth similarity. The reconstruction loss is calculated based on the training data pairs, the multimodal data time series, and the global vector. The global vector is decoupled to obtain the attribute vector, and the threshold penalty loss is calculated based on the attribute vector and the training data pair. The threshold penalty loss, reconstruction loss, mean squared error loss, and orthogonal constraint loss are weighted and summed to obtain the total loss. The model parameters with the minimum total loss are then obtained to obtain the optimal feature extraction model.

3. The similarity measurement method for complex landscapes as described in claim 2, characterized in that, The multimodal time series data matrix is ​​subjected to forward encoding and backward encoding respectively to obtain an output vector. A concatenation vector is obtained from the output vector. A normalized weight is obtained from the concatenation vector. A time series feature is obtained from the normalized weight and the concatenation vector. The time series features at different time steps form a time series sequence. The multimodal time-series data matrix is ​​encoded using forward LSTM to obtain the forward output vector; The multimodal time-series data matrix is ​​then encoded using inverse LSTM to obtain an inverse output vector. The forward output vector and the reverse output vector are concatenated to obtain the concatenated vector; Calculate the attention score for each concatenated vector and normalize the attention score to obtain the normalized weight; The time series features are obtained based on the normalized weights and the concatenated vectors, and the time series features at different time steps form a time series sequence.

4. A similarity measurement method for complex landscapes as described in claim 2, characterized in that, The step of obtaining the global vector based on the time series includes: The time series is processed by multi-head self-attention to obtain projection features, and multi-head splicing features are obtained based on the projection features. Add residual connections to the multi-head splicing features to obtain residual features, and normalize the residual features to obtain normalized residual features; Calculate the attention score of the residual normalized feature, and obtain the feature weight based on the attention score; The global vector is obtained based on the feature weights and attention scores.

5. A similarity measurement method for complex landscapes as described in claim 1, characterized in that, The attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector, and is expressed as follows: in, For attribute loss, For the first The attribute threshold corresponding to each attribute vector. For the first A vector of attributes The total number of attributes. For the first The values ​​of the attribute vector.

6. A similarity measurement method for complex landscapes as described in claim 1, characterized in that, The comprehensive similarity is obtained based on the Mahalanobis distance, attribute entropy, mutual information, and change similarity. The similarity loss is then obtained based on this comprehensive similarity, expressed as follows: in, , , as well as As weight, The Mahalanobis distance, For the similarity of changes, For mutual information, This is the attribute entropy.

7. A similarity measurement system for complex landscapes, characterized in that, include: The data acquisition module is used to acquire a landscape training dataset for the garden landscape, and to acquire multimodal time series data of the garden landscape based on the landscape training dataset. The training data includes data on changes in scenery, landscape interest data, distribution data of landscape elements, changes in the opening and closing of landscape features, and landscape aesthetic data. The sequence of the degree of scenery variation is used as the data on changes in scenery variation, and the sequence of the degree of scenery variation is represented as follows: ,in, For the rate of change of landscape type, This indicates the landscape type label corresponding to the visitor's location at time t. The landscape type is determined by a predefined set of landscape element categories. The information entropy represents the distribution of landscape elements at time t, where c is the index of the landscape element category. This represents the probability of the c-th type of landscape element appearing at time t, used to characterize the degree of diversity in the local landscape composition. For the surprise of the transition matrix, To establish a stable segment length distribution, t represents time; the aesthetic theme sequence is used as landscape aesthetic data, and the aesthetic theme sequence is represented as follows: Extracting topic tags through natural language processing. This refers to the aesthetic theme tags associated with the visitor's location or scenic node at time t. The aesthetic themes are obtained through natural language processing of landscape text descriptions, historical and cultural introductions, or expert-annotated texts. The set of theme tags is determined by a pre-constructed aesthetic theme thesaurus. () is a one-hot encoding function used to map discrete aesthetic topic labels into binary vectors of fixed dimensions, the dimension of which is equal to the total number of aesthetic topic categories. The dimension corresponding to the current topic is set to 1, and the other dimensions are set to 0. The data conversion module is used to convert the multimodal data time series into time series vectors, and multiple time series vectors form a multimodal time series data matrix; The feature extraction model training module is used to set up the feature extraction model, train the feature extraction model based on the multimodal time series data matrix to obtain the optimal feature extraction model, and obtain the optimal global vector based on the optimal feature extraction model and the multimodal time series data matrix. The decoupling mapping module is used to decouple and map the optimal global vector to obtain the attribute vector; The similarity model optimization module is used to input the attribute vector and training dataset into the similarity model to optimize the similarity model and obtain the optimal similarity model. The actual data acquisition module is used to acquire actual landscape data for different garden landscapes; The similarity calculation module is used to input the actual landscape data into the optimal feature extraction model to obtain the actual global vectors of different garden landscapes, and input the actual global vectors of different garden landscapes into the optimal similarity model to obtain the similarity of different landscapes. The step of inputting the attribute vector and training dataset into the similarity model to optimize the similarity model and obtain the optimal similarity model includes: The attribute loss is calculated based on the attribute vector and the attribute threshold corresponding to the attribute vector. Mahalanobis distance, mutual information, and attribute entropy are calculated based on the attribute vectors of different landscapes. The Euclidean distance is calculated based on the attribute vectors of different landscapes, and the variation similarity is obtained based on the Euclidean distance and the attribute vectors of different landscapes. A comprehensive similarity is obtained based on the Mahalanobis distance, attribute entropy, mutual information, and change similarity, and a similarity loss is obtained based on the comprehensive similarity. Obtain training data pairs based on the training dataset, and obtain reconstruction loss based on the training data pairs; Based on the reconstruction loss, similarity loss, and attribute loss, the total similarity loss is obtained. The similarity model is then optimized based on the total similarity loss to obtain an optimized similarity model.

8. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 6.

Images