Method and system for intelligent combination and scoring of clothing style elements

Abstract

The application provides a clothing style element intelligent combination and scoring method and system, and relates to the technical field of artificial intelligence, which comprises the following steps: mapping clothing style elements to a vector space, constructing a semantic correlation matrix for feature enhancement, constructing a bidirectional relationship graph based on the enhanced features and enumerating connected subgraphs as candidate schemes, calculating internal consistency scores and structural stability scores, and solving a combination coordination score to screen out a target style combination scheme. The application can effectively improve the coordination and aesthetic degree of clothing style element combination, and realizes personalized clothing matching recommendation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for intelligent combination and scoring of clothing style elements. Background Technology

[0002] With social development and rising consumption levels, people's aesthetic demands for clothing are becoming increasingly diversified. The rational combination of clothing style elements has become a crucial consideration in modern clothing design and personal styling. Clothing style elements refer to the various components that constitute the overall style of clothing, such as color, material, texture, cut, and silhouette. The intelligent combination and scoring of these elements is of great significance for improving clothing design efficiency, optimizing e-commerce recommendation systems, and assisting consumers in making personalized styling decisions.

[0003] Currently, the combination of clothing style elements mainly relies on the designer's professional experience and subjective judgment, lacking objective quantitative evaluation methods. Existing methods for matching clothing style elements still ignore the semantic connections and contextual dependencies between elements, resulting in a lack of holistic consideration in the combination evaluation results. They use simple rule matching or shallow feature extraction, which cannot effectively express the deep semantic features of style elements. They also lack the ability to model complex style element combination relationships in a refined manner and lack a multi-dimensional comprehensive evaluation mechanism, making it difficult to simultaneously take into account the internal consistency and structural stability of the combination scheme. Summary of the Invention

[0004] This invention provides a method and system for intelligent combination and scoring of clothing style elements, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides a method for intelligent combination and scoring of clothing style elements, comprising:

[0006] Multiple style elements of the clothing object to be processed are obtained and mapped to a unified vector space through multi-dimensional feature encoding to obtain an initial feature vector. The semantic similarity between the initial feature vectors is calculated and a semantic association matrix between style elements is constructed. Based on the semantic association matrix, the initial feature vector is enhanced with context awareness. By weighted aggregation of style element features with high semantic association, the context-enhanced feature vector of each style element is obtained.

[0007] Based on the context-enhanced feature vector, a bidirectional relationship graph between style elements is constructed, and subgraphs are enumerated in the bidirectional relationship graph. A set of connected subgraphs is generated by combining depth-first traversal, and the set of style elements corresponding to each connected subgraph is used as a candidate style combination scheme.

[0008] The cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme is calculated to obtain the element similarity matrix. The matrix norm of the element similarity matrix is ​​calculated and the internal consistency score of the scheme is determined. The topological characteristic parameters of the candidate style combination scheme in the bidirectional relationship graph are extracted and the structural stability score of the scheme is calculated.

[0009] The combination coordination score of each candidate style combination scheme is obtained by solving the internal consistency score and the structural stability score of the scheme. The target style combination scheme is then selected based on the combination coordination score.

[0010] In one alternative implementation,

[0011] The process involves obtaining multiple style elements of the clothing object to be processed and mapping them to a unified vector space using multi-dimensional feature encoding to obtain initial feature vectors. Calculating the semantic similarity between these initial feature vectors and constructing a semantic association matrix between style elements includes:

[0012] The system receives visual input data of the clothing object to be processed and performs semantic segmentation to obtain multiple semantic regions. It calculates the color histogram distribution and texture gradient response corresponding to each semantic region and determines the style attribute category of the semantic region. It performs spatial connectivity detection and boundary fusion on semantic regions with the same style attribute category to obtain multiple style elements.

[0013] Extract the curvature change sequence of the shape contour, the main hue vector of the color space, and the spectral energy distribution of the surface texture corresponding to each style element. Then, normalize and concatenate the numerical range to obtain multi-dimensional original features. Perform linear projection on the multi-dimensional original features to obtain intermediate feature representations. Apply a nonlinear activation function to the intermediate feature representations and perform dimensionality reduction projection to obtain the initial feature vectors of each style element.

[0014] The initial feature vectors are paired up, and the inner product of each pair of initial feature vectors is calculated to obtain the original similarity value. The scaled similarity value is calculated based on the number of dimensions of the initial feature vectors. The semantic similarity between each style element pair is obtained by performing an exponential function transformation on the scaled similarity value.

[0015] A two-dimensional matrix is ​​constructed based on style elements, and the semantic similarity is filled into the two-dimensional matrix according to the row index and column index of the style element pairs. The upper triangular cells and lower triangular cells are extracted by traversing the two-dimensional matrix to construct the semantic association matrix between style elements.

[0016] In one alternative implementation,

[0017] Based on the semantic association matrix, the initial feature vector is enhanced with context-aware features. By weighted aggregation of style element features with high semantic relevance, the context-enhanced feature vectors of each style element are obtained, including:

[0018] Extract the semantic association degree row vector corresponding to each style element from the semantic association matrix and sort them in descending order according to the numerical value to obtain the sorted association degree sequence. Calculate the cumulative summation curve of the sorted association degree sequence and determine the cutoff position where the cumulative sum reaches a preset proportion of the total. Extract the style element index corresponding to each element before the cutoff position and mark it as the high association neighborhood set of the current style element.

[0019] The element values ​​at the corresponding positions are extracted from the semantic relevance row vector based on the highly correlated neighborhood set as neighborhood relevance weights. The corresponding initial feature vectors are extracted from the highly correlated neighborhood set to obtain the neighborhood feature sequence. The neighborhood relevance weights are mapped by an exponential function and then summed to obtain a normalized denominator. The attention weights are then calculated based on the normalized denominator.

[0020] The attention weights and the neighborhood feature sequence are summed element-wise to obtain weighted aggregate features. The variance of each dimension of the weighted aggregate features is calculated and its reciprocal is taken to obtain the uncertainty weight vector. The variance weighted features are obtained by solving based on the weighted aggregate features and the uncertainty weight vector.

[0021] The variance-weighted features are added element-wise to the initial feature vector of the current style element to obtain the fused feature vector. The absolute value of each dimension of the fused feature vector is taken and the mean value is calculated to obtain the scaling reference value. The context-enhanced feature vector is obtained based on the fused feature vector and the scaling reference value.

[0022] In one alternative implementation,

[0023] Based on the context-enhanced feature vectors, a bidirectional relationship graph between style elements is constructed, and subgraphs are enumerated in the bidirectional relationship graph. A set of connected subgraphs is generated by combining depth-first traversal. The set of style elements corresponding to each connected subgraph is used as a candidate style combination scheme, including:

[0024] Obtain the context-enhanced feature vectors corresponding to the style elements and perform joint training using graph embedding learning methods to obtain the structure embedding vectors. Combine the style elements in pairs to obtain style element pairs. Extract the structure embedding vectors corresponding to the style element pairs and calculate the distance metric. If the distance metric is less than a preset distance threshold, establish a bidirectional edge between the two style elements. Repeat the judgment and construction of bidirectional edges to obtain a bidirectional relationship graph.

[0025] The bidirectional relationship graph is decomposed to obtain a set of low-dimensional embedding representations of nodes. The range of the number of nodes in the subgraph is set based on the distance distribution between nodes in the low-dimensional embedding representation set. All node combinations with the number of nodes within the range of the number of nodes are enumerated to determine the enumerated subgraph and construct the enumerated subgraph set.

[0026] For each node in the enumerated subgraph, the node access flag is initialized to unvisited, and a node is randomly selected as the starting node. All enumerated subgraphs are traversed, and a depth-first traversal is performed to verify and obtain a set of connected subgraphs. The current connected subgraph corresponding to the starting node is determined. All adjacent nodes of the starting node in the enumerated subgraph are obtained and added to the current connected subgraph to determine the access flag. If it is unvisited, the adjacent node is used as the starting node, and the process of obtaining adjacent nodes and determining the flag is repeated until all adjacent nodes are processed. It is determined whether the number of nodes in the current connected subgraph is equal to the number of nodes in the enumerated subgraph. If so, the style element corresponding to the current connected subgraph is extracted to obtain a candidate style combination scheme.

[0027] In one alternative implementation,

[0028] Calculate the cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme to obtain an element similarity matrix. Calculate the matrix norm of the element similarity matrix and determine the scheme's internal consistency score, including:

[0029] Extract all style elements and their corresponding context-enhanced feature vectors contained in the candidate style combination scheme. Map the context-enhanced feature vectors to a unified metric space through vector projection transformation to obtain a metric space vector. Combine all style elements in pairs to obtain style element pairs within the scheme. Calculate the cosine similarity value between the metric space vectors corresponding to the style element pairs within the scheme. Construct an element similarity matrix based on the cosine similarity value and the index position of the style element in the candidate style combination scheme.

[0030] The element similarity matrix is ​​decomposed into singular value sequence and singular vector matrix. The Frobenius norm of each singular value in the singular value sequence is calculated. The singular vector corresponding to the largest singular value in the singular value sequence is extracted and the variance of each component is determined to obtain the dominant direction dispersion. The Frobenius norm and the dominant direction dispersion are weighted and combined to obtain the matrix structure stability index.

[0031] The arithmetic mean of all cosine similarity values ​​in the element similarity matrix is ​​calculated as the average similarity of the schemes. The number of element pairs in the element similarity matrix with cosine similarity values ​​greater than a preset similarity threshold is counted. The proportion of high similarity is determined by combining the total number of style element pairs in the scheme. The internal consistency score of the scheme is obtained by combining the average similarity of the schemes and the matrix structure stability index.

[0032] In one alternative implementation,

[0033] Extracting the topological characteristic parameters of the candidate style combination schemes from the bidirectional relationship graph and calculating the scheme structural stability score includes:

[0034] Traverse the nodes within the candidate style combination scheme and obtain the corresponding adjacent nodes. Count the number of adjacent nodes belonging to the current candidate style combination scheme. Determine the node cohesion coefficient based on the total number of adjacent nodes. Calculate the arithmetic mean of the cohesion coefficients of all nodes in the current candidate style combination scheme as the scheme average cohesion coefficient. Count the number of nodes in the current candidate style combination scheme whose node cohesion coefficient is greater than the preset cohesion threshold. Calculate the proportion of high cohesion nodes based on the total number of nodes in the current candidate style combination scheme.

[0035] For any two nodes in the current candidate style combination scheme, search for the shortest path connecting the two nodes in the bidirectional relationship graph, determine the number of edges contained in the shortest path to obtain the distance between the nodes, calculate the arithmetic mean of the distances between all node pairs in the current candidate style combination scheme as the average path length of the scheme, count the number of node pairs in the current candidate style combination scheme whose distance is less than a preset distance threshold, and calculate the ratio of the number of node pairs to the total number of node pairs in the current candidate style combination scheme as the tight connection ratio.

[0036] The structural stability score of the scheme is obtained by solving the ratio of high cohesion nodes and the ratio of tight connections.

[0037] In one alternative implementation,

[0038] Based on the internal consistency score and structural stability score of the proposed scheme, the combination coordination score of each candidate style combination scheme is obtained. Based on the combination coordination score, the target style combination schemes are selected, including:

[0039] Obtain the internal consistency score and structural stability score of each candidate style combination scheme and calculate the Pearson correlation coefficient. Determine the adaptive weight coefficient based on the Pearson correlation coefficient. Based on the adaptive weight coefficient, perform weighted fusion of the internal consistency score and structural stability score of the scheme to obtain the preliminary coordination score.

[0040] Extract the relative position of the internal consistency score of the current candidate style combination scheme among all candidate style combination schemes to determine the relative ranking of internal consistency; extract the relative position of the structural stability score of the current candidate style combination scheme among all candidate style combination schemes to determine the relative ranking of structural stability.

[0041] The ranking deviation is obtained by calculating the degree of deviation between the relative ranking of internal consistency and the relative ranking of structural stability. A coordination penalty factor is constructed based on the ranking deviation, and the combined coordination score is obtained by combining the preliminary coordination score.

[0042] Iterate through all candidate style combination schemes and calculate the combination coordination score. Sort all candidate style combination schemes in descending order according to the combination coordination score and select the first candidate style combination scheme as the target style combination scheme for output.

[0043] A second aspect of this invention provides an intelligent combination and scoring system for clothing style elements, comprising:

[0044] The first unit is used to obtain multiple style elements of the clothing object to be processed and map them to a unified vector space through multi-dimensional feature encoding to obtain an initial feature vector. It calculates the semantic similarity between the initial feature vectors and constructs a semantic association matrix between style elements. Based on the semantic association matrix, it performs context-aware feature enhancement on the initial feature vectors and obtains the context-enhanced feature vector of each style element by weighted aggregation of style element features with high semantic association.

[0045] The second unit is used to construct a bidirectional relationship graph between style elements based on the context-enhanced feature vector and to enumerate subgraphs of the bidirectional relationship graph. It also uses a depth-first traversal to generate a set of connected subgraphs and uses the set of style elements corresponding to each connected subgraph as a candidate style combination scheme.

[0046] The third unit is used to calculate the cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme to obtain the element similarity matrix, calculate the matrix norm of the element similarity matrix and determine the internal consistency score of the scheme, extract the topological characteristic parameters of the candidate style combination scheme in the bidirectional relationship graph and calculate the scheme structure stability score.

[0047] The fourth unit is used to calculate the combination coordination score of each candidate style combination scheme based on the internal consistency score and the structural stability score of the scheme, and to select the target style combination scheme based on the combination coordination score.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0050] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0051] In this invention, context-aware feature enhancement of style elements is achieved through multi-dimensional feature encoding and semantic association matrix construction, making feature representation more accurate and semantically rich. A bidirectional relationship graph and subgraph enumeration method are used to combine clothing style elements, combined with depth-first traversal to generate a set of connected subgraphs, effectively solving the problems of high computational complexity and low combination efficiency in traditional exhaustive methods. This achieves more efficient generation of style combination schemes. By calculating the internal consistency score and structural stability score of the scheme, not only the semantic similarity between style elements is considered, but also the topological characteristics of the combination structure, making the scoring results more objective, comprehensive, and reliable. Based on the combination coordination score, intelligent screening of style combination schemes is achieved, which can automatically identify and recommend the best style element combination scheme, significantly improving the professionalism and intelligence level of clothing matching. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the intelligent combination and scoring method for clothing style elements according to an embodiment of the present invention.

[0053] Figure 2 This is a flowchart illustrating the stability evaluation of the style combination structure of the intelligent combination and scoring method for clothing style elements in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0056] Figure 1 This is a flowchart illustrating the intelligent combination and scoring method for clothing style elements according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] Multiple style elements of the clothing object to be processed are obtained and mapped to a unified vector space through multi-dimensional feature encoding to obtain an initial feature vector. The semantic similarity between the initial feature vectors is calculated and a semantic association matrix between style elements is constructed. Based on the semantic association matrix, the initial feature vector is enhanced with context awareness. By weighted aggregation of style element features with high semantic association, the context-enhanced feature vector of each style element is obtained.

[0058] Based on the context-enhanced feature vector, a bidirectional relationship graph between style elements is constructed, and subgraphs are enumerated in the bidirectional relationship graph. A set of connected subgraphs is generated by combining depth-first traversal, and the set of style elements corresponding to each connected subgraph is used as a candidate style combination scheme.

[0059] The cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme is calculated to obtain the element similarity matrix. The matrix norm of the element similarity matrix is ​​calculated and the internal consistency score of the scheme is determined. The topological characteristic parameters of the candidate style combination scheme in the bidirectional relationship graph are extracted and the structural stability score of the scheme is calculated.

[0060] The combination coordination score of each candidate style combination scheme is obtained by solving the internal consistency score and the structural stability score of the scheme. The target style combination scheme is then selected based on the combination coordination score.

[0061] In one alternative implementation,

[0062] The process involves obtaining multiple style elements of the clothing object to be processed and mapping them to a unified vector space using multi-dimensional feature encoding to obtain initial feature vectors. Calculating the semantic similarity between these initial feature vectors and constructing a semantic association matrix between style elements includes:

[0063] The system receives visual input data of the clothing object to be processed and performs semantic segmentation to obtain multiple semantic regions. It calculates the color histogram distribution and texture gradient response corresponding to each semantic region and determines the style attribute category of the semantic region. It performs spatial connectivity detection and boundary fusion on semantic regions with the same style attribute category to obtain multiple style elements.

[0064] Extract the curvature change sequence of the shape contour, the main hue vector of the color space, and the spectral energy distribution of the surface texture corresponding to each style element. Then, normalize and concatenate the numerical range to obtain multi-dimensional original features. Perform linear projection on the multi-dimensional original features to obtain intermediate feature representations. Apply a nonlinear activation function to the intermediate feature representations and perform dimensionality reduction projection to obtain the initial feature vectors of each style element.

[0065] The initial feature vectors are paired up, and the inner product of each pair of initial feature vectors is calculated to obtain the original similarity value. The scaled similarity value is calculated based on the number of dimensions of the initial feature vectors. The semantic similarity between each style element pair is obtained by performing an exponential function transformation on the scaled similarity value.

[0066] A two-dimensional matrix is ​​constructed based on style elements, and the semantic similarity is filled into the two-dimensional matrix according to the row index and column index of the style element pairs. The upper triangular cells and lower triangular cells are extracted by traversing the two-dimensional matrix to construct the semantic association matrix between style elements.

[0067] A deep convolutional neural network is used to perform semantic segmentation on the input clothing image. For example, for an image of a dress, semantic segmentation can obtain multiple semantic regions, such as the collar region, cuff region, skirt body region, and hem region. A color histogram distribution is calculated for each semantic region, dividing the RGB color space into 64 color intervals, and the pixel proportion within each interval is statistically analyzed to obtain a 64-dimensional color histogram. The texture gradient response of the semantic region is calculated by performing convolution operations on the region using Sobel operators in the horizontal and vertical directions to obtain gradient strength and direction information. The gradient direction is divided into 8 intervals, and the pixel proportion within each interval is statistically analyzed to form an 8-dimensional texture feature vector.

[0068] Based on color histograms and texture features, style attribute categories are determined for each semantic region. Predefined style attribute categories include minimalist, retro, sporty, ethnic, and elegant styles. The determination process uses a support vector machine (SVM) classifier. The input is a concatenated vector of the color histogram and texture features, and the output is the probability distribution of the style attribute category for the semantic region. For example, the hem area of ​​a dress might have a probability value of 0.75 for elegant style and 0.25 for retro style after classification. The style attribute category for that region is taken as the elegant style corresponding to the highest probability.

[0069] Spatial connectivity detection and boundary fusion are performed on semantic regions with the same style attribute category to obtain multiple style elements. Spatial connectivity detection uses a region growing algorithm, starting with a single pixel and gradually merging adjacent pixels with the same style attribute category until no further expansion is possible. Boundary fusion uses morphological closing operations, filling the gaps between regions with the same style attribute category and a spacing of less than 10 pixels with circular structural elements of a radius of 5 pixels. For example, if both the skirt body and hem areas are in the "elegant" style, they can be merged into a single style element through spatial connectivity detection and boundary fusion.

[0070] After style element identification, multidimensional features of each style element are extracted, including the curvature change sequence of the shape contour. A sampling point is taken every 10 pixels along the style element boundary, and the angle formed by three adjacent sampling points is calculated to obtain a 32-dimensional vector describing the contour curvature change. The dominant hue vector in the color space is extracted, and K-means clustering is applied to the style element region with a cluster size of 5, resulting in 5 dominant hues and their proportions, forming a 10-dimensional dominant hue vector. The spectral energy distribution of the surface texture is extracted, and a two-dimensional discrete Fourier transform is applied to the style element region, dividing the frequency domain into high-frequency, mid-frequency, and low-frequency intervals. The energy proportion of each interval is calculated to form a 3-dimensional spectral energy vector.

[0071] The extracted multi-dimensional original features are numerically normalized, mapping each dimension's value to between 0 and 1, and concatenated to form a 45-dimensional original feature vector. A linear projection is then performed on the original feature vector, multiplying it by a 45×128 weight matrix to obtain a 128-dimensional intermediate feature representation. A modified linear unit activation function is applied to the intermediate feature representation, setting negative values ​​to 0 and retaining positive values ​​unchanged. This is then dimensionality-reduced by projection using a 128×64 weight matrix, resulting in a 64-dimensional initial feature vector.

[0072] Semantic similarity is calculated by pairwise pairing of the obtained initial feature vectors. Assume there is a shirt and a pair of pants, with initial feature vectors A and B being 64-dimensional vectors respectively. The inner product of vectors A and B is calculated to obtain the original similarity value. If the original similarity value is 38.5, and considering the dimensionality of the initial feature vectors (64), the scaled similarity value is calculated. Dividing 38.5 by 64 yields 0.602. An exponential function transformation is then applied to the scaled similarity value using an exponential function with the natural logarithm as the base, resulting in a semantic similarity of 1.826.

[0073] A two-dimensional matrix is ​​constructed based on style elements and filled with semantic similarity. Assuming the identified style elements include five elements: collar, cuffs, body, waistband, and hem, a 5×5 two-dimensional matrix is ​​constructed. Semantic similarity is filled based on the row and column indices of the style element pairs. For example, if the semantic similarity between the body and hem is 1.826, the semantic similarity is filled into the corresponding position in the matrix. The two-dimensional matrix is ​​traversed, and the values ​​of the upper and lower triangular cells are extracted to construct the semantic association matrix between style elements.

[0074] In this embodiment, by performing semantic segmentation on clothing images and comprehensively introducing color distribution and texture response at the semantic region level for style attribute determination, the influence of background interference and non-style regions on the analysis results is effectively reduced, improving the robustness and consistency of style recognition. By performing spatial connectivity detection and boundary fusion on semantic regions with consistent style attributes, structurally complete and semantically consistent style elements are formed, enhancing the stability and structural rationality of style expression and effectively avoiding the problem of incomplete style description caused by region fragmentation. By performing linear mapping and nonlinear transformation on multidimensional features to form a compact initial feature vector, the feature dimension and redundancy are reduced while maintaining style discrimination information, improving the numerical stability and computational efficiency of subsequent similarity calculations, and overcoming the noise amplification and scale inconsistency problems caused by direct comparison of high-dimensional features.

[0075] In one alternative implementation,

[0076] Based on the semantic association matrix, the initial feature vector is enhanced with context-aware features. By weighted aggregation of style element features with high semantic relevance, the context-enhanced feature vectors of each style element are obtained, including:

[0077] Extract the semantic association degree row vector corresponding to each style element from the semantic association matrix and sort them in descending order according to the numerical value to obtain the sorted association degree sequence. Calculate the cumulative summation curve of the sorted association degree sequence and determine the cutoff position where the cumulative sum reaches a preset proportion of the total. Extract the style element index corresponding to each element before the cutoff position and mark it as the high association neighborhood set of the current style element.

[0078] The element values ​​at the corresponding positions are extracted from the semantic relevance row vector based on the highly correlated neighborhood set as neighborhood relevance weights. The corresponding initial feature vectors are extracted from the highly correlated neighborhood set to obtain the neighborhood feature sequence. The neighborhood relevance weights are mapped by an exponential function and then summed to obtain a normalized denominator. The attention weights are then calculated based on the normalized denominator.

[0079] The attention weights and the neighborhood feature sequence are summed element-wise to obtain weighted aggregate features. The variance of each dimension of the weighted aggregate features is calculated and its reciprocal is taken to obtain the uncertainty weight vector. The variance weighted features are obtained by solving based on the weighted aggregate features and the uncertainty weight vector.

[0080] The variance-weighted features are added element-wise to the initial feature vector of the current style element to obtain the fused feature vector. The absolute value of each dimension of the fused feature vector is taken and the mean value is calculated to obtain the scaling reference value. The context-enhanced feature vector is obtained based on the fused feature vector and the scaling reference value.

[0081] The association degree information of each style element is extracted from a pre-constructed semantic association matrix. For example, a 5x5 semantic association matrix has been established for the five identified clothing style elements: neckline, cuff, waist, skirt body, and hem. Taking the neckline element as an example, the first row is extracted from the semantic association matrix as the semantic association degree row vector of the neckline, with values ​​of 0, 1.826, 1.523, 1.247, and 0.985, representing the semantic association degree between the neckline and itself, cuff, waist, skirt body, and hem, respectively. This row vector is sorted from largest to smallest, resulting in a sorted association degree sequence of 1.826, 1.523, 1.247, 0.985, and 0, corresponding to the element indices of cuff, waist, skirt body, hem, and neckline itself, respectively.

[0082] Calculate the cumulative sum curve of the sorted correlation sequence to determine the range of highly correlated neighborhoods. The cumulative sum sequence of the neckline elements is 1.826, 3.349, 4.596, 5.581, 5.581. Set a preset ratio threshold of 0.8, and calculate the truncation threshold as 5.581 multiplied by 0.8, which equals 4.465. Find the first value exceeding the threshold of 4.465 in the cumulative sum sequence, which is 4.596, located at the 3rd position in the sequence. Determine the truncation position as 3, and extract the element indices before this position as cuff, waist, and skirt body. These three elements constitute the highly correlated neighborhood set of the neckline elements.

[0083] Using a defined set of highly correlated neighborhoods, the element values ​​at corresponding positions are extracted from the original semantic correlation row vector as neighborhood correlation weights. The highly correlated neighborhood set for the neckline element includes the cuff, waist, and skirt, with corresponding correlation weights of 1.826, 1.523, and 1.247, respectively. Based on the element indices of the highly correlated neighborhood set, the corresponding initial feature vectors are extracted to form a neighborhood feature sequence. The initial feature vectors for the cuff, waist, and skirt elements are 64-dimensional vectors, with the values ​​of the first three dimensions as follows: cuff: 0.752, 0.614, 0.893; waist: 0.685, 0.573, 0.824; skirt: 0.631, 0.549, 0.778.

[0084] An exponential function is applied to the neighborhood association weights to amplify the weight differences. Using an exponential function with the natural logarithm as the base, 1.826, 1.523, and 1.247 are mapped to 6.209, 4.586, and 3.480, respectively. The sum of the mapped weights is used as the normalized denominator, which has a value of 14.275. Dividing each mapped weight by the normalized denominator yields the attention weights of 0.435, 0.321, and 0.244.

[0085] The attention weights are summed element-wise with the neighborhood feature sequence to obtain the weighted aggregated features. For example, for the first dimension of the feature vector, 0.752 multiplied by 0.435 plus 0.685 multiplied by 0.321 plus 0.631 multiplied by 0.244 is calculated, resulting in 0.702. The second dimension is calculated as 0.614 multiplied by 0.435 plus 0.573 multiplied by 0.321 plus 0.549 multiplied by 0.244, resulting in 0.582. The third dimension is calculated as 0.893 multiplied by 0.435 plus 0.824 multiplied by 0.321 plus 0.778 multiplied by 0.244, resulting in 0.838. This process is repeated for all 64 dimensions of the weighted aggregated features.

[0086] The dispersion of the neighborhood feature sequence across each dimension is calculated to evaluate the reliability of the weighted aggregated features. Taking the first dimension as an example, the differences between the feature values ​​of the three neighborhood elements (0.752, 0.685, 0.631) and the weighted aggregated value (0.702) are 0.050, -0.017, and -0.071, respectively. The squares of these differences are 0.0025, 0.0003, and 0.0050. Multiplying the squared differences by the attention weights and summing the results yields 0.0024. Taking the square root gives the standard deviation of 0.049, and taking its reciprocal gives the uncertainty weight of 20.408, indicating high consistency and strong reliability in the first dimension. The same calculations are performed on the other dimensions to obtain the complete uncertainty weight vector.

[0087] Multiplying the weighted aggregate features by the uncertainty weight vector yields the variance-weighted features. The variance-weighted feature for the first dimension is 0.702 multiplied by 20.408, equaling 14.326. The second dimension is 0.582 multiplied by 23.256, equaling 13.535. The third dimension is 0.838 multiplied by 17.857, equaling 14.963. Variance-weighted features for all dimensions are calculated in this manner.

[0088] The variance-weighted features are added to the initial feature vector of the current style element to obtain the fused feature vector. Assuming the first three dimensions of the initial feature vector for the neckline element are 0.724, 0.631, and 0.865, then the first three dimensions of the fused feature vector are 14.326 + 0.724 = 15.050, 13.535 + 0.631 = 14.166, and 14.963 + 0.865 = 15.828. The fused feature values ​​for all dimensions are calculated in this manner.

[0089] To control the numerical range of the fused features, the fused feature vector is normalized. The average absolute value of each dimension of the fused feature vector is calculated as the scaling benchmark. Assume the average absolute value of the 64-dimensional fused feature vector is 12.855. Dividing each dimension of the fused feature vector by this scaling benchmark yields the context-enhanced feature vector. The values ​​for the first three dimensions are 15.050 divided by 12.855 equals 1.171, 14.166 divided by 12.855 equals 1.102, and 15.828 divided by 12.855 equals 1.231. This process is repeated for all dimensions to calculate the normalized values, resulting in the complete context-enhanced feature vector.

[0090] The same context enhancement process is applied to other style elements in the garment. Taking the cuff element as an example, the second row of the semantic association matrix is ​​extracted as its semantic association degree row vector, with values ​​of 1.826, 0, 1.342, 1.169, and 0.914, representing the semantic association degree between the cuff and the neckline, the cuff itself, the waist, the skirt body, and the hem, respectively. After sorting, cumulative summation, truncation position determination, neighborhood association weight extraction, attention weight calculation, weighted aggregation feature calculation, uncertainty weight solution, variance weighted feature calculation, fusion feature acquisition, and feature normalization, the context-enhanced feature vector of the cuff element is obtained.

[0091] The same context enhancement process is also applied to the waist, skirt body, and hem elements, ultimately yielding context-enhanced feature vectors for all style elements.

[0092] In this embodiment, by extracting and sorting the correlation information of each style element from the semantic correlation matrix, and adaptively determining the high correlation neighborhood range based on the cumulative ratio, each style element establishes a contextual relationship only with its most representative associated object, avoiding the introduction of irrelevant information and the spread of semantic noise, thus improving the pertinence and accuracy of style correlation modeling. By introducing the correlation within the high correlation neighborhood into the exponential mapping and normalizing it to form attention weights, the correlation strength is dynamically adjusted in a continuous and distinguishable manner during the aggregation process, significantly enhancing the expressive power of context modeling. By reverse-weighting the variance of each dimension of the weighted aggregation features, the feature dimensions with higher stability and smaller fluctuations have a greater influence in the fusion process, improving the robustness of feature representation to noise and abnormal correlations in multi-style element interaction scenarios, and improving the problem that context fusion is susceptible to local anomalies.

[0093] In one alternative implementation,

[0094] Based on the context-enhanced feature vectors, a bidirectional relationship graph between style elements is constructed, and subgraphs are enumerated in the bidirectional relationship graph. A set of connected subgraphs is generated by combining depth-first traversal. The set of style elements corresponding to each connected subgraph is used as a candidate style combination scheme, including:

[0095] Obtain the context-enhanced feature vectors corresponding to the style elements and perform joint training using graph embedding learning methods to obtain the structure embedding vectors. Combine the style elements in pairs to obtain style element pairs. Extract the structure embedding vectors corresponding to the style element pairs and calculate the distance metric. If the distance metric is less than a preset distance threshold, establish a bidirectional edge between the two style elements. Repeat the judgment and construction of bidirectional edges to obtain a bidirectional relationship graph.

[0096] The bidirectional relationship graph is decomposed to obtain a set of low-dimensional embedding representations of nodes. The range of the number of nodes in the subgraph is set based on the distance distribution between nodes in the low-dimensional embedding representation set. All node combinations with the number of nodes within the range of the number of nodes are enumerated to determine the enumerated subgraph and construct the enumerated subgraph set.

[0097] For each node in the enumerated subgraph, the node access flag is initialized to unvisited, and a node is randomly selected as the starting node. All enumerated subgraphs are traversed, and a depth-first traversal is performed to verify and obtain a set of connected subgraphs. The current connected subgraph corresponding to the starting node is determined. All adjacent nodes of the starting node in the enumerated subgraph are obtained and added to the current connected subgraph to determine the access flag. If it is unvisited, the adjacent node is used as the starting node, and the process of obtaining adjacent nodes and determining the flag is repeated until all adjacent nodes are processed. It is determined whether the number of nodes in the current connected subgraph is equal to the number of nodes in the enumerated subgraph. If so, the style element corresponding to the current connected subgraph is extracted to obtain a candidate style combination scheme.

[0098] The context-enhanced feature vectors corresponding to each style element are obtained. Taking a dress containing five style elements—neckline, cuffs, waist, skirt body, and hem—as an example, the context-enhanced feature vectors for each of the five style elements are 64-dimensional vectors. These feature vectors are then jointly trained using a graph embedding learning method to obtain a structural embedding vector. The graph embedding learning method employs random walk and skip training, preserving the original feature space topology while compressing the high-dimensional features into a 32-dimensional structural embedding vector.

[0099] Style elements are paired to obtain style element pairs. For five style elements, a total of 10 style element pairs are formed: neckline and cuff, neckline and waist, neckline and skirt body, neckline and hem, cuff and waist, cuff and skirt body, cuff and hem, waist and skirt body, waist and hem, and skirt body and hem. The structural embedding vectors corresponding to each style element pair are extracted, and the distance metric between them is calculated. Euclidean distance is used; for the neckline and cuff pair, the Euclidean distance between the corresponding structural embedding vectors is 0.563. A preset distance threshold of 0.85 is set. Since 0.563 is less than 0.85, a bidirectional edge is established between the neckline and cuff.

[0100] Similarly, the distance metrics for other style element pairs are calculated: neckline to waist is 0.726, neckline to skirt body is 0.912, neckline to hem is 1.137, cuff to waist is 0.804, cuff to skirt body is 0.892, cuff to hem is 1.065, waist to skirt body is 0.638, waist to hem is 0.743, and skirt body to hem is 0.571. Compared with the preset threshold of 0.85, the distance metrics for neckline to skirt body, neckline to hem, cuff to skirt body, and cuff to hem are greater than 0.85, so no connection is established; the distance metrics for neckline to cuff, neckline to waist, cuff to waist, waist to skirt body, waist to hem, and skirt body to hem are less than 0.85, so a two-way edge is established. After completing the judgment and connection construction of all style element pairs, a bidirectional relationship graph containing 5 nodes and 6 bidirectional edges is obtained.

[0101] The constructed bidirectional relationship graph is subjected to graph decomposition to obtain a set of low-dimensional embedding representations of the nodes. Graph decomposition is achieved based on the eigenvalue decomposition of the graph Laplacian matrix, compressing the graph structure information into 8-dimensional vectors. For the collar node, its low-dimensional embedding representation is an 8-dimensional vector with values ​​of 0.412, 0.376, 0.298, 0.215, 0.183, 0.142, 0.094, and 0.068. The low-dimensional embedding representations of other nodes are generated similarly, forming a set of low-dimensional embedding representations containing 5 nodes, each corresponding to an 8-dimensional vector.

[0102] The range of the number of nodes in the subgraph is set based on the distance distribution between nodes in the low-dimensional embedding representation set. The Euclidean distance between any two nodes in the low-dimensional embedding space is calculated to obtain the distance distribution. The characteristics of the distance distribution are analyzed, and the range of the number of nodes in the subgraph is set to 3 to 5.

[0103] Enumerate all node combinations with a number of nodes within a specified range. For the case with 3 nodes, selecting 3 nodes from 5 nodes results in 10 combinations. For the case with 4 nodes, there are 5 combinations. For the case with 5 nodes, there is only 1 combination. This results in a total of 16 potential subgraphs, forming the enumerated subgraph set.

[0104] For each subgraph in the enumerated subgraph set, its connectivity is verified to determine valid style combination schemes. Taking the node set collar, cuff, and waist as an example, the access markers of the three nodes are first initialized to unvisited. The collar is randomly selected as the starting node, marked as visited, and the current connected subgraph corresponding to the starting node is determined, initially containing only the collar node. All adjacent nodes of the collar in the enumerated subgraph are obtained, including the cuff and waist. The cuff is added to the current connected subgraph, marked as visited, and the search continues with the cuff as the new starting node. The cuff's adjacent nodes include the collar and waist. Since the collar has been visited, the waist is added to the current connected subgraph and marked as visited. At this point, all nodes are marked as visited, and the connected subgraph contains the three nodes collar, cuff, and waist, which is equal to the number of nodes in the enumerated subgraph. Therefore, the enumerated subgraph is connected, and the corresponding style element combination of collar, cuff, and waist is extracted as a candidate style combination scheme.

[0105] Similar connectivity verification was performed on other enumerated subgraphs. For the node sets collar, cuff, and hem, a depth-first traversal revealed that only a connected subgraph containing the collar and cuff nodes could be formed. The number of nodes was less than the three nodes in the enumerated subgraph, therefore it did not constitute a valid candidate style combination. Through connectivity verification on all 16 enumerated subgraphs, seven connected subgraphs were obtained, corresponding to seven candidate style combinations: collar, cuff, waist; collar, waist, skirt body; waist, skirt body, hem; cuff, waist, skirt body; collar, cuff, waist, skirt body; waist, skirt body, hem; collar, cuff, waist, skirt body, hem.

[0106] In this embodiment, by performing graph embedding joint learning on context-enhanced feature vectors, the high-dimensional semantic relationships between style elements are mapped into an embedding space that maintains adjacency structure and relative distance, improving the consistency and stability of style relationship modeling at the global scale. By constructing a bidirectional relationship graph based on the embedding space distance threshold, the associations between style elements are expressed in a symmetrical and explicit graph structure, filtering out weakly related or accidentally similar element pairs, ensuring the inherent consistency required for style combination, and significantly reducing the probability of unreasonable combinations. By performing low-dimensional embedding decomposition on the bidirectional relationship graph and adaptively determining the subgraph size range by combining the distance distribution between nodes, the subsequent combination search is carried out in a semantically dense and reasonably sized region, improving the efficiency and targeting of the combination generation process.

[0107] In one alternative implementation,

[0108] Calculate the cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme to obtain an element similarity matrix. Calculate the matrix norm of the element similarity matrix and determine the scheme's internal consistency score, including:

[0109] Extract all style elements and their corresponding context-enhanced feature vectors contained in the candidate style combination scheme. Map the context-enhanced feature vectors to a unified metric space through vector projection transformation to obtain a metric space vector. Combine all style elements in pairs to obtain style element pairs within the scheme. Calculate the cosine similarity value between the metric space vectors corresponding to the style element pairs within the scheme. Construct an element similarity matrix based on the cosine similarity value and the index position of the style element in the candidate style combination scheme.

[0110] The element similarity matrix is ​​decomposed into singular value sequence and singular vector matrix. The Frobenius norm of each singular value in the singular value sequence is calculated. The singular vector corresponding to the largest singular value in the singular value sequence is extracted and the variance of each component is determined to obtain the dominant direction dispersion. The Frobenius norm and the dominant direction dispersion are weighted and combined to obtain the matrix structure stability index.

[0111] The arithmetic mean of all cosine similarity values ​​in the element similarity matrix is ​​calculated as the average similarity of the schemes. The number of element pairs in the element similarity matrix with cosine similarity values ​​greater than a preset similarity threshold is counted. The proportion of high similarity is determined by combining the total number of style element pairs in the scheme. The internal consistency score of the scheme is obtained by combining the average similarity of the schemes and the matrix structure stability index.

[0112] Extract all style elements and their corresponding context-enhanced feature vectors from the candidate style combination scheme. Taking a candidate style combination scheme composed of three style elements—collar, cuff, and waist—as an example, extract the context-enhanced feature vectors of these three style elements. Each vector is a 64-dimensional vector. The first five components of the feature vector for the collar element are 1.171, 1.102, 1.231, 0.905, and 0.762; the first five components of the feature vector for the cuff element are 1.058, 0.975, 1.196, 0.873, and 0.824; and the first five components of the feature vector for the waist element are 0.986, 1.032, 1.142, 0.919, and 0.793.

[0113] The context-enhanced feature vectors are mapped to a unified metric space through vector projection transformation, resulting in a metric space vector. The vector projection transformation uses a linear mapping method to convert the 64-dimensional feature vectors into a 32-dimensional metric space vector. A pre-trained projection matrix is ​​then multiplied by the context-enhanced feature vectors to obtain the metric space vector. The first five components of the metric space vector for the neckline element are 0.896, 0.823, 0.751, 0.687, and 0.624; for the cuff element, they are 0.857, 0.792, 0.736, 0.675, and 0.612; and for the waist element, they are 0.827, 0.768, 0.712, 0.693, and 0.635.

[0114] All style elements are paired to obtain style element pairs within the scheme. For the three elements of neckline, cuff, and waist, three style element pairs are obtained: neckline and cuff, neckline and waist, and cuff and waist. The cosine similarity value between the corresponding metric space vectors of each style element pair is calculated. The cosine similarity is calculated by dividing the dot product of the two vectors by the product of their magnitudes. The cosine similarity value between neckline and cuff is 0.983, between neckline and waist is 0.941, and between cuff and waist is 0.956.

[0115] An element similarity matrix is ​​constructed based on the cosine similarity value and the index position of the style element in the candidate style combination scheme. For the case of three style elements, a 3x3 element similarity matrix is ​​constructed. The values ​​on the diagonal of the matrix represent the similarity between an element and itself, all of which are 1.000. The values ​​off-diagonal represent the cosine similarity values ​​between different elements. The element similarity matrix composed of three elements—collar, cuff, and waist—is as follows: the first row is 1.000, 0.983, and 0.941; the second row is 0.983, 1.000, and 0.956; and the third row is 0.941, 0.956, and 1.000.

[0116] Singular value decomposition (SVD) is performed on the element similarity matrix to obtain a sequence of singular values ​​and a singular vector matrix. SVD decomposes the element similarity matrix into the product of three matrices: a left singular vector matrix, a diagonal singular value matrix, and the transpose of a right singular vector matrix. Performing SVD on the aforementioned 3x3 element similarity matrix yields the singular value sequence 2.911, 0.058, and 0.031. The first column of the left singular vector matrix, corresponding to the singular vectors with the largest singular values, contains the following values: 0.577, 0.581, and 0.574.

[0117] Calculate the Frobenius norm for each singular value in the singular value sequence. The Frobenius norm is obtained by calculating the square root of the sum of squares of the singular values, which is 2.912. Extract the singular vector corresponding to the largest singular value in the singular value sequence, and calculate the variance of each component of the singular vector to obtain the dominant direction dispersion. The singular vectors corresponding to the largest singular value 2.911 are 0.577, 0.581, and 0.574, and the variance of these three components yields a dominant direction dispersion of 0.0000123. The smaller the dominant direction dispersion, the more stable the matrix structure.

[0118] The matrix structure stability index is obtained by weighting the Frobenius norm and the dispersion of the dominant direction. The combination method is to divide the Frobenius norm by the dispersion of the dominant direction and multiply by a scaling factor of 0.01, resulting in a matrix structure stability index value of 23.673.

[0119] The arithmetic mean of all cosine similarity values ​​in the element similarity matrix is ​​calculated as the average similarity of the schemes. The average of the off-diagonal cosine similarity values ​​(0.983, 0.941, 0.956) is calculated, yielding an average similarity of 0.960. The number of element pairs in the element similarity matrix with cosine similarity values ​​greater than a preset similarity threshold is counted, and this, combined with the total number of style element pairs within the scheme, determines the high similarity percentage. With a preset similarity threshold of 0.950, two element pairs in the element similarity matrix have cosine similarity values ​​greater than 0.950: 0.983 for the neckline and cuff, and 0.956 for the cuff and waist. The total number of style element pairs within the scheme is 3, and the high similarity percentage is 2 divided by 3, which equals 0.667.

[0120] By combining the average similarity of the schemes and the matrix structure stability index, the internal consistency score of the schemes is obtained. The calculation method is to multiply the average similarity of the schemes by 0.7 and add the normalized value of the matrix structure stability index by 0.3, resulting in an internal consistency score of 0.835. The higher the internal consistency score, the more harmonious the style elements in the candidate style combination schemes are.

[0121] The same method was used to calculate the internal consistency score for other candidate style combinations. For example, for a candidate scheme composed of three style elements—waist, skirt body, and hem—the calculated element similarity matrix was as follows: the first row is 1.000, 0.972, and 0.913; the second row is 0.972, 1.000, and 0.965; and the third row is 0.913, 0.965, and 1.000. After singular value decomposition, matrix structure stability index calculation, average scheme similarity calculation, and high similarity ratio statistics, the internal consistency score was 0.817.

[0122] In this embodiment, by uniformly mapping the contextual enhancement features of each style element to the same metric space, the similarity calculation between different style elements is established on a consistent scale and semantic benchmark, avoiding the similarity distortion problem caused by differences in feature sources or distributions, and improving the comparability and accuracy of similarity assessment within the scheme. By constructing an element similarity matrix based on cosine similarity, the interrelationships between all style element pairs within the scheme are characterized as a whole in matrix structure, enhancing the analytical capability for complex multi-element combinations. By performing singular value decomposition on the element similarity matrix and introducing Frobenius norm and dominant direction dispersion as joint indicators, the concentration of relationships and directional consistency within the scheme are measured simultaneously, which can effectively distinguish combination schemes with similar surface average similarity but significantly different internal structural stability, overcoming the problem of insufficient discrimination capability caused by relying solely on mean similarity evaluation.

[0123] In one alternative implementation,

[0124] Extracting the topological characteristic parameters of the candidate style combination schemes from the bidirectional relationship graph and calculating the scheme structural stability score includes:

[0125] Traverse the nodes within the candidate style combination scheme and obtain the corresponding adjacent nodes. Count the number of adjacent nodes belonging to the current candidate style combination scheme. Determine the node cohesion coefficient based on the total number of adjacent nodes. Calculate the arithmetic mean of the cohesion coefficients of all nodes in the current candidate style combination scheme as the scheme average cohesion coefficient. Count the number of nodes in the current candidate style combination scheme whose node cohesion coefficient is greater than the preset cohesion threshold. Calculate the proportion of high cohesion nodes based on the total number of nodes in the current candidate style combination scheme.

[0126] For any two nodes in the current candidate style combination scheme, search for the shortest path connecting the two nodes in the bidirectional relationship graph, determine the number of edges contained in the shortest path to obtain the distance between the nodes, calculate the arithmetic mean of the distances between all node pairs in the current candidate style combination scheme as the average path length of the scheme, count the number of node pairs in the current candidate style combination scheme whose distance is less than a preset distance threshold, and calculate the ratio of the number of node pairs to the total number of node pairs in the current candidate style combination scheme as the tight connection ratio.

[0127] The structural stability score of the scheme is obtained by solving the ratio of high cohesion nodes and the ratio of tight connections.

[0128] Traverse the nodes within the candidate style combination scheme and obtain the corresponding adjacent nodes. Taking a candidate style combination scheme composed of three style elements—collar, cuff, and waist—as an example, in the bidirectional relationship graph constructed above, the adjacent nodes of the collar include the cuff and waist; the adjacent nodes of the cuff include the collar and waist; and the adjacent nodes of the waist include the collar, cuff, skirt body, and hem.

[0129] The number of adjacent nodes belonging to the current candidate style combination scheme is counted, and the node cohesion coefficient is determined by combining this number with the total number of adjacent nodes. The node cohesion coefficient represents the degree of connection between a node and other nodes within the candidate style combination scheme, and is calculated by dividing the number of adjacent nodes belonging to the scheme by the total number of adjacent nodes. For the collar node, its total number of adjacent nodes is 2, all of which belong to the current candidate style combination scheme, so the cohesion coefficient is 2 divided by 2, which equals 1.000; for the cuff node, its total number of adjacent nodes is 2, all of which belong to the current candidate style combination scheme, so the cohesion coefficient is 2 divided by 2, which equals 1.000; for the waist node, its total number of adjacent nodes is 4, of which 2 belong to the current candidate style combination scheme, so the cohesion coefficient is 2 divided by 4, which equals 0.500.

[0130] The arithmetic mean of the cohesion coefficients of all nodes within the current candidate style combination scheme is calculated as the average cohesion coefficient of the scheme. For example, the cohesion coefficients of the neckline, cuff, and waist nodes are 1.000, 1.000, and 0.500, respectively, and the arithmetic mean yields an average cohesion coefficient of 0.833. The number of nodes within the current candidate style combination scheme whose cohesion coefficient is greater than a preset cohesion threshold is counted, and combined with the total number of nodes in the current candidate style combination scheme, the proportion of high-cohesion nodes is calculated. Setting the preset cohesion threshold to 0.800, the cohesion coefficients of the neckline and cuff are greater than the preset threshold, while the cohesion coefficient of the waist is less than the preset threshold. Therefore, the number of high-cohesion nodes is 2, the total number of nodes is 3, and the proportion of high-cohesion nodes is 2 divided by 3, which equals 0.667.

[0131] For any two nodes within the current candidate style combination scheme, search for the shortest path connecting the two nodes in the bidirectional relationship graph, and determine the distance between the nodes by the number of edges contained in the shortest path. The shortest path search is implemented using a breadth-first search algorithm. For the neckline and cuff, they are directly connected, the shortest path contains 1 edge, and the distance between the nodes is 1; for the neckline and waist, they are directly connected, the shortest path contains 1 edge, and the distance between the nodes is 1; for the cuff and waist, they are directly connected, the shortest path contains 1 edge, and the distance between the nodes is 1.

[0132] Calculate the arithmetic mean of the distances between all node pairs within the current candidate style combination scheme, as the average path length of the scheme. The distances between the three node pairs (collar and cuff, collar and waist, cuff and waist) are all 1, and the arithmetic mean yields an average path length of 1.000. Count the number of node pairs within the current candidate style combination scheme whose distance is less than a preset distance threshold, and calculate the ratio of this number to the total number of node pairs in the current candidate style combination scheme, as the tight connection ratio. Set the preset distance threshold to 2. Since the distances between the three node pairs are all 1, which is less than the preset distance threshold, the number of tightly connected node pairs is 3. The total number of node pairs is 3, and the tight connection ratio is 3 divided by 3, which equals 1.000.

[0133] The structural stability score of the proposed scheme is obtained by calculating the proportion of highly cohesive nodes and the proportion of tightly connected nodes. The calculation method is to multiply the proportion of highly cohesive nodes by 0.6 and add the proportion of tightly connected nodes by 0.4, resulting in a structural stability score of 0.800. A higher structural stability score indicates better stability of the candidate style combination scheme in the graph structure, and a more compliant combination of style elements with design specifications.

[0134] The same method was used to calculate the structural stability score for other candidate style combinations. Taking a candidate scheme composed of three style elements—waist, skirt body, and hem—as an example, the adjacent nodes of the waist include the neckline, cuffs, skirt body, and hem. Among these, two nodes—the skirt body and the hem—belong to the current scheme, with a cohesion coefficient of 2 divided by 4, equal to 0.500. The adjacent nodes of the skirt body include the waist and the hem, both belonging to the current scheme, with a cohesion coefficient of 2 divided by 2, equal to 1.000. The adjacent nodes of the hem include the waist and the skirt body, both belonging to the current scheme, with a cohesion coefficient of 2 divided by 2, equal to 1.000. The average cohesion coefficient of the scheme is 0.833, and the proportion of high-cohesion nodes is 0.667.

[0135] Regarding the distances between nodes, the distances between the three pairs of nodes—waist and skirt body, waist and hem, and skirt body and hem—are all 1. The average path length of the scheme is 1.000, and the proportion of tightly connected nodes is 1.000. The calculated structural stability score of this scheme is 0.800.

[0136] Candidate designs comprised of four style elements: neckline, cuffs, waist, and skirt body. The neckline's adjacent nodes include the cuffs and waist, all within the current design, with a cohesion coefficient of 1.000. The cuffs' adjacent nodes also include the neckline and waist, all within the current design, with a cohesion coefficient of 1.000. The waist's adjacent nodes include the neckline, cuffs, skirt body, and hem, with three nodes (neckline, cuffs, and skirt body) within the current design, resulting in a cohesion coefficient of 0.750. The skirt body's adjacent nodes include the waist and hem, with only one node (waist) within the current design, resulting in a cohesion coefficient of 0.500. The average cohesion coefficient for the designs is 0.813, and the proportion of high-cohesion nodes is 0.750.

[0137] Regarding the distances between nodes, the distances between the four pairs of nodes—neckline and cuff, neckline and waist, cuff and waist, and waist and skirt—are all 1; the distance between the neckline and skirt is 2; and the distance between the cuff and skirt is 2. The average path length of the scheme is 1.333. Six out of the six pairs of nodes have a distance less than the preset distance threshold of 2, and the proportion of tightly connected nodes is 1.000. The final calculated structural stability score for this scheme is 0.850.

[0138] Candidate designs comprised of five style elements—neckline, cuffs, waist, skirt body, and hem—were scored. The cohesion coefficients for the neckline, cuffs, waist, skirt body, and hem were all 0.500, 1.000, 0.667, and 0.667, respectively. The average cohesion coefficient was 0.667, and the proportion of highly cohesive nodes was 0.400. For the distance between ten pairs of nodes, the average path length was 1.600, and the proportion of tightly connected nodes was 0.800. The calculated structural stability score for the current design was 0.560.

[0139] In this embodiment, by statistically analyzing the adjacency relationships of each node within the candidate style combination scheme and calculating the node cohesion coefficient, the local connectivity tightness of each style element in the combination is quantitatively characterized. This accurately reflects the degree of integration between a single style element and the overall scheme, avoiding the problem of ignoring the quality of local structures by judging solely based on the number of elements or simple similarity. By statistically analyzing the node cohesion coefficients within the scheme and introducing a high cohesion node ratio index, the evaluation results not only focus on the average level but also highlight whether most nodes have good structural support, thus improving the discriminative ability of structural stability assessment. By introducing shortest path distance analysis into the bidirectional relationship graph, the global reachability and connection efficiency between style elements within the scheme are quantified, revealing the collaborative connectivity characteristics within the scheme more comprehensively.

[0140] Figure 2 This is a flowchart illustrating the stability evaluation of the style combination structure of the intelligent combination and scoring method for clothing style elements in an embodiment of the present invention.

[0141] In one alternative implementation,

[0142] Based on the internal consistency score and structural stability score of the proposed scheme, the combination coordination score of each candidate style combination scheme is obtained. Based on the combination coordination score, the target style combination schemes are selected, including:

[0143] Obtain the internal consistency score and structural stability score of each candidate style combination scheme and calculate the Pearson correlation coefficient. Determine the adaptive weight coefficient based on the Pearson correlation coefficient. Based on the adaptive weight coefficient, perform weighted fusion of the internal consistency score and structural stability score of the scheme to obtain the preliminary coordination score.

[0144] Extract the relative position of the internal consistency score of the current candidate style combination scheme among all candidate style combination schemes to determine the relative ranking of internal consistency; extract the relative position of the structural stability score of the current candidate style combination scheme among all candidate style combination schemes to determine the relative ranking of structural stability.

[0145] The ranking deviation is obtained by calculating the degree of deviation between the relative ranking of internal consistency and the relative ranking of structural stability. A coordination penalty factor is constructed based on the ranking deviation, and the combined coordination score is obtained by combining the preliminary coordination score.

[0146] Iterate through all candidate style combination schemes and calculate the combination coordination score. Sort all candidate style combination schemes in descending order according to the combination coordination score and select the first candidate style combination scheme as the target style combination scheme for output.

[0147] For each candidate style combination, the internal consistency score and structural stability score were obtained, and the Pearson correlation coefficient was calculated. Taking the aforementioned seven candidate style combinations as examples, the internal consistency scores were 0.835, 0.817, 0.768, 0.793, 0.842, 0.781, and 0.752, respectively, and the structural stability scores were 0.800, 0.800, 0.750, 0.770, 0.850, 0.765, and 0.560, respectively. The Pearson correlation coefficient between the two sets of scores was calculated to be 0.678, indicating a strong positive correlation between the two types of scores.

[0148] Adaptive weighting coefficients were determined based on the Pearson correlation coefficient. When the Pearson correlation coefficient is positive and greater than 0.5, it indicates that the two types of scores are relatively consistent, and the internal consistency score and structural stability score contribute similarly to the overall evaluation; in this case, a relatively balanced weighting configuration is adopted. When the Pearson correlation coefficient is close to 0, it indicates that the two types of scores are independent and each provides different evaluation dimensions; in this case, equal weights are adopted. When the Pearson correlation coefficient is negative, it indicates that the two types of scores conflict, and the weights need to be adjusted according to the specific application scenario. Based on a Pearson correlation coefficient of 0.678, the adaptive weighting coefficient for the internal consistency score was determined to be 0.6, and the adaptive weighting coefficient for the structural stability score was determined to be 0.4.

[0149] A preliminary coordination score is obtained by weighting and fusing the internal consistency score and structural stability score of the proposed schemes based on adaptive weighting coefficients. For example, for the first candidate style combination scheme, the internal consistency score is 0.835 and the structural stability score is 0.800. The weighted fusion calculation is 0.835 multiplied by 0.6 plus 0.800 multiplied by 0.4, resulting in a preliminary coordination score of 0.821. Similarly, the preliminary coordination scores for the other six candidate style combinations are calculated to be 0.810, 0.761, 0.784, 0.845, 0.775, and 0.676, respectively.

[0150] Extract the relative position of the internal consistency score of the current candidate style combination among all candidate style combinations to determine the relative ranking of internal consistency. The internal consistency scores of the seven candidate style combinations are sorted from highest to lowest as follows: 0.842, 0.835, 0.817, 0.793, 0.781, 0.768, and 0.752, corresponding to rankings of 1, 2, 3, 4, 5, 6, and 7. The first candidate style combination has an internal consistency score of 0.835, ranking 2nd, and its relative ranking is 2 divided by 7, equal to 0.286. The second candidate style combination has an internal consistency score of 0.817, ranking 3rd, and its relative ranking is 3 divided by 7, equal to 0.429.

[0151] Extract the relative position of the structural stability score of the current candidate style combination among all candidate style combinations to determine the relative ranking of structural stability. The structural stability scores of the seven candidate style combinations are sorted from highest to lowest as 0.850, 0.800, 0.800, 0.770, 0.765, 0.750, and 0.560, corresponding to rankings of 1, 2.5, 2.5, 4, 5, 6, and 7. Since the first and second candidate style combinations both have a structural stability score of 0.800, ranking second, an average ranking of 2.5 is adopted. The relative ranking of structural stability for the first candidate style combination is 2.5 divided by 7, which equals 0.357. The relative ranking of structural stability for the second candidate style combination is 2.5 divided by 7, which equals 0.357.

[0152] The ranking deviation is calculated by determining the degree of deviation between the relative ranking of internal consistency and the relative ranking of structural stability. The ranking deviation is derived from the absolute difference between the two relative rankings. For the first candidate style combination, the relative ranking of internal consistency is 0.286, and the relative ranking of structural stability is 0.357. The ranking deviation is the absolute difference between 0.286 and 0.357, which equals 0.071. For the second candidate style combination, the relative ranking of internal consistency is 0.429, and the relative ranking of structural stability is 0.357. The ranking deviation is the absolute difference between 0.429 and 0.357, which equals 0.072.

[0153] A consistency penalty factor is constructed based on ranking deviation. The consistency penalty factor is calculated by subtracting the ranking deviation multiplied by a penalty coefficient from 1, with the penalty coefficient set to 2. The ranking deviation of the first candidate style combination is 0.071, and the consistency penalty factor is 1 minus 0.071 multiplied by 2, equaling 0.858. The ranking deviation of the second candidate style combination is 0.072, and the consistency penalty factor is 1 minus 0.072 multiplied by 2, equaling 0.856.

[0154] The combined coordination score is obtained by combining the initial coordination score and the coordination penalty factor. The combined coordination score is calculated by multiplying the initial coordination score by the coordination penalty factor. The initial coordination score of the first candidate style combination is 0.821, the coordination penalty factor is 0.858, and the combined coordination score is 0.821 multiplied by 0.858, which equals 0.704. The initial coordination score of the second candidate style combination is 0.810, the coordination penalty factor is 0.856, and the combined coordination score is 0.810 multiplied by 0.856, which equals 0.693.

[0155] The combination coordination score is calculated by iterating through all candidate style combinations. The third candidate style combination has an internal consistency score of 0.768, a structural stability score of 0.750, a preliminary coordination score of 0.761, a relative ranking of internal consistency of 0.857, a relative ranking of structural stability of 0.857, a ranking deviation of 0.000, a coordination penalty factor of 1.000, and a combination coordination score of 0.761. The fourth candidate style combination has a combination coordination score of 0.659. The fifth candidate style combination has a combination coordination score of 0.719. The sixth candidate style combination has a combination coordination score of 0.689. The seventh candidate style combination has a combination coordination score of 0.498.

[0156] All candidate style combinations are sorted in descending order based on their combinational harmony scores. The sorted results are: the fifth candidate style combination, the first candidate style combination, the third candidate style combination, the second candidate style combination, the sixth candidate style combination, the fourth candidate style combination, and the seventh candidate style combination, with corresponding combinational harmony scores of 0.719, 0.704, 0.761, 0.693, 0.689, 0.659, and 0.498, respectively. The fifth candidate style combination with the highest combinational harmony score is selected as the target style combination.

[0157] The fifth candidate style combination scheme includes four style elements: neckline, cuffs, waist, and skirt. It has a high score in internal consistency and structural stability, and the rankings of the two scores are relatively consistent. This indicates that the scheme not only conforms to aesthetic and design rules in the combination of style elements, but also has good structural stability. It is a highly coordinated clothing style element combination scheme.

[0158] In this embodiment, by introducing the Pearson correlation coefficient between the internal consistency score and the structural stability score of the scheme and adaptively determining the weight accordingly, different evaluation indicators can dynamically adjust their influence based on their statistical correlation during the fusion process. This avoids the indicator bias problem caused by using fixed weights or empirical weights, thereby improving the ability of the comprehensive evaluation results to reflect the actual quality of the style combination. By weighting and fusing the two types of scores to form a preliminary coordination score, the style consistency level and structural reliability of the combined scheme are considered simultaneously. Compared with the method of focusing on only a single evaluation dimension, this effectively reduces the risk of misselecting schemes with high consistency but loose structure or structural stability but fragmented style. By introducing the ranking deviation as a coordination penalty factor, schemes with serious imbalances in internal consistency and structural stability are suppressed, thereby improving the overall coordination and reliability of the target style combination in practical applications.

[0159] A second aspect of this invention provides an intelligent combination and scoring system for clothing style elements, comprising:

[0160] The first unit is used to obtain multiple style elements of the clothing object to be processed and map them to a unified vector space through multi-dimensional feature encoding to obtain an initial feature vector. It calculates the semantic similarity between the initial feature vectors and constructs a semantic association matrix between style elements. Based on the semantic association matrix, it performs context-aware feature enhancement on the initial feature vectors and obtains the context-enhanced feature vector of each style element by weighted aggregation of style element features with high semantic association.

[0161] The second unit is used to construct a bidirectional relationship graph between style elements based on the context-enhanced feature vector and to enumerate subgraphs of the bidirectional relationship graph. It also uses a depth-first traversal to generate a set of connected subgraphs and uses the set of style elements corresponding to each connected subgraph as a candidate style combination scheme.

[0162] The third unit is used to calculate the cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme to obtain the element similarity matrix, calculate the matrix norm of the element similarity matrix and determine the internal consistency score of the scheme, extract the topological characteristic parameters of the candidate style combination scheme in the bidirectional relationship graph and calculate the scheme structure stability score.

[0163] The fourth unit is used to calculate the combination coordination score of each candidate style combination scheme based on the internal consistency score and the structural stability score of the scheme, and to select the target style combination scheme based on the combination coordination score.

[0164] A third aspect of the present invention provides an electronic device, comprising:

[0165] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0166] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0167] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent combination and scoring of clothing style elements, characterized in that, include: Multiple style elements of the clothing object to be processed are obtained and mapped to a unified vector space through multi-dimensional feature encoding to obtain initial feature vectors. The semantic similarity between the initial feature vectors is calculated and a semantic association matrix between style elements is constructed. Based on the semantic association matrix, context-aware feature enhancement is performed on the initial feature vectors. By weighted aggregation of style element features with high semantic relevance, the context-enhanced feature vectors of each style element are obtained, specifically including: Extract the semantic association degree row vector corresponding to each style element from the semantic association matrix and sort them in descending order according to the numerical value to obtain the sorted association degree sequence. Calculate the cumulative summation curve of the sorted association degree sequence and determine the cutoff position where the cumulative sum reaches a preset proportion of the total. Extract the style element index corresponding to each element before the cutoff position and mark it as the high association neighborhood set of the current style element. The element values ​​at the corresponding positions are extracted from the semantic relevance row vector based on the highly correlated neighborhood set as neighborhood relevance weights. The corresponding initial feature vectors are extracted from the highly correlated neighborhood set to obtain the neighborhood feature sequence. The neighborhood relevance weights are mapped by an exponential function and then summed to obtain a normalized denominator. The attention weights are then calculated based on the normalized denominator. The attention weights and the neighborhood feature sequence are summed element-wise to obtain weighted aggregate features. The standard deviations of each dimension of the weighted aggregate features are calculated and their reciprocals are taken to obtain the uncertainty weight vector. The variance weighted features are obtained by solving based on the weighted aggregate features and the uncertainty weight vector. The variance-weighted features are added element-wise to the initial feature vector of the current style element to obtain the fused feature vector. The absolute value of each dimension of the fused feature vector is taken and the mean value is calculated to obtain the scaling reference value. The context enhancement feature vector is obtained based on the fused feature vector and the scaling reference value. The process of obtaining multiple style elements of the clothing object to be processed includes receiving visual input data of the clothing object to be processed and performing semantic segmentation to obtain multiple semantic regions, calculating the color histogram distribution and texture gradient response corresponding to each semantic region and determining the style attribute category of the semantic region, and performing spatial connectivity detection and boundary fusion on semantic regions with the same style attribute category to obtain multiple style elements. Based on the context-enhanced feature vector, a bidirectional relationship graph between style elements is constructed, and subgraphs are enumerated in the bidirectional relationship graph. A set of connected subgraphs is generated by combining depth-first traversal, and the set of style elements corresponding to each connected subgraph is used as a candidate style combination scheme. The cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme is calculated to obtain an element similarity matrix. The matrix norm of the element similarity matrix is ​​calculated, and the internal consistency score of the scheme is determined. The topological characteristic parameters of the candidate style combination scheme in the bidirectional relationship graph are extracted, and the structural stability score of the scheme is calculated. Specifically, this includes: Extract all style elements and their corresponding context-enhanced feature vectors contained in the candidate style combination scheme. Map the context-enhanced feature vectors to a unified metric space through vector projection transformation to obtain a metric space vector. Combine all style elements in pairs to obtain style element pairs within the scheme. Calculate the cosine similarity value between the metric space vectors corresponding to the style element pairs within the scheme. Construct an element similarity matrix based on the cosine similarity value and the index position of the style element in the candidate style combination scheme. The element similarity matrix is ​​decomposed into singular value sequence and singular vector matrix. The Frobenius norm of each singular value in the singular value sequence is calculated. The singular vector corresponding to the largest singular value in the singular value sequence is extracted and the variance of each component is determined to obtain the dominant direction dispersion. The Frobenius norm and the dominant direction dispersion are weighted and combined to obtain the matrix structure stability index. The arithmetic mean of all cosine similarity values ​​in the element similarity matrix is ​​calculated as the average similarity of the scheme. The number of element pairs in the element similarity matrix with cosine similarity values ​​greater than a preset similarity threshold is counted. The proportion of high similarity is determined by combining the total number of style element pairs in the scheme. The internal consistency score of the scheme is obtained by combining the average similarity of the scheme and the matrix structure stability index. Traverse the nodes within the candidate style combination scheme and obtain the corresponding adjacent nodes. Count the number of adjacent nodes belonging to the current candidate style combination scheme. Determine the node cohesion coefficient based on the total number of adjacent nodes. Calculate the arithmetic mean of the cohesion coefficients of all nodes in the current candidate style combination scheme as the scheme average cohesion coefficient. Count the number of nodes in the current candidate style combination scheme whose node cohesion coefficient is greater than the preset cohesion threshold. Calculate the proportion of high cohesion nodes based on the total number of nodes in the current candidate style combination scheme. For any two nodes in the current candidate style combination scheme, search for the shortest path connecting the two nodes in the bidirectional relationship graph, determine the number of edges contained in the shortest path to obtain the distance between the nodes, calculate the arithmetic mean of the distances between all node pairs in the current candidate style combination scheme as the average path length of the scheme, count the number of node pairs in the current candidate style combination scheme whose distance is less than a preset distance threshold, and calculate the ratio of the number of node pairs to the total number of node pairs in the current candidate style combination scheme as the tight connection ratio. The structural stability score of the scheme is obtained by solving based on the proportion of high cohesion nodes and the proportion of tight connections. Based on the internal consistency score and structural stability score of the proposed scheme, the combination coordination score of each candidate style combination scheme is obtained. The target style combination scheme is then selected based on the combination coordination score, specifically including: Obtain the internal consistency score and structural stability score of each candidate style combination scheme and calculate the Pearson correlation coefficient. Determine the adaptive weight coefficient based on the Pearson correlation coefficient. Based on the adaptive weight coefficient, perform weighted fusion of the internal consistency score and structural stability score of the scheme to obtain the preliminary coordination score. Extract the relative position of the internal consistency score of the current candidate style combination scheme among all candidate style combination schemes to determine the relative ranking of internal consistency; extract the relative position of the structural stability score of the current candidate style combination scheme among all candidate style combination schemes to determine the relative ranking of structural stability. The ranking deviation is obtained by calculating the degree of deviation between the relative ranking of internal consistency and the relative ranking of structural stability. A coordination penalty factor is constructed based on the ranking deviation, and the combined coordination score is obtained by combining the preliminary coordination score. Iterate through all candidate style combination schemes and calculate the combination coordination score. Sort all candidate style combination schemes in descending order according to the combination coordination score and select the first candidate style combination scheme as the target style combination scheme for output.

2. The method according to claim 1, characterized in that, Initial feature vectors are obtained by mapping multi-dimensional feature encoding to a unified vector space. The semantic similarity between these initial feature vectors and the construction of a semantic association matrix between style elements include: Extract the curvature change sequence of the shape contour, the main hue vector of the color space, and the spectral energy distribution of the surface texture corresponding to each style element. Then, normalize and concatenate the numerical range to obtain multi-dimensional original features. Perform linear projection on the multi-dimensional original features to obtain intermediate feature representations. Apply a nonlinear activation function to the intermediate feature representations and perform dimensionality reduction projection to obtain the initial feature vectors of each style element. The initial feature vectors are paired up, and the inner product of each pair of initial feature vectors is calculated to obtain the original similarity value. The scaled similarity value is calculated based on the number of dimensions of the initial feature vectors. The semantic similarity between each style element pair is obtained by performing an exponential function transformation on the scaled similarity value. A two-dimensional matrix is ​​constructed based on style elements, and the semantic similarity is filled into the two-dimensional matrix according to the row index and column index of the style element pairs. The upper triangular cells and lower triangular cells are extracted by traversing the two-dimensional matrix to construct the semantic association matrix between style elements.

3. The method according to claim 1, characterized in that, Based on the context-enhanced feature vectors, a bidirectional relationship graph between style elements is constructed, and subgraphs are enumerated in the bidirectional relationship graph. A set of connected subgraphs is generated by combining depth-first traversal. The set of style elements corresponding to each connected subgraph is used as a candidate style combination scheme, including: Obtain the context-enhanced feature vectors corresponding to style elements and perform joint training using graph embedding learning methods to obtain structural embedding vectors. Combine the style elements in pairs to obtain style element pairs. Extract the structural embedding vectors corresponding to the style element pairs and calculate the distance metric. If the distance metric is less than a preset distance threshold, establish a bidirectional edge between the two style elements. Repeat the judgment and construction of bidirectional edges to obtain a bidirectional relationship graph. The bidirectional relationship graph is decomposed to obtain a set of low-dimensional embedding representations of nodes. The range of the number of nodes in the subgraph is set based on the distance distribution between nodes in the low-dimensional embedding representation set. All node combinations with the number of nodes within the range of the number of nodes are enumerated to determine the enumerated subgraph and construct the enumerated subgraph set. For each node in the enumerated subgraph, the node access flag is initialized to unvisited, and a node is randomly selected as the starting node. All enumerated subgraphs are traversed, and a depth-first traversal is performed to verify and obtain a set of connected subgraphs. The current connected subgraph corresponding to the starting node is determined. All adjacent nodes of the starting node in the enumerated subgraph are obtained and added to the current connected subgraph to determine the access flag. If it is unvisited, the adjacent node is used as the starting node, and the process of obtaining adjacent nodes and determining the flag is repeated until all adjacent nodes are processed. It is determined whether the number of nodes in the current connected subgraph is equal to the number of nodes in the enumerated subgraph. If so, the style element corresponding to the current connected subgraph is extracted to obtain a candidate style combination scheme.

4. A smart combination and scoring system for clothing style elements, used to implement the method of any one of claims 1-3, characterized in that it includes: The first unit is used to obtain multiple style elements of the clothing object to be processed and map them to a unified vector space through multi-dimensional feature encoding to obtain an initial feature vector. It calculates the semantic similarity between the initial feature vectors and constructs a semantic association matrix between style elements. Based on the semantic association matrix, it performs context-aware feature enhancement on the initial feature vectors and obtains the context-enhanced feature vector of each style element by weighted aggregation of style element features with high semantic association. The second unit is used to construct a bidirectional relationship graph between style elements based on the context-enhanced feature vector and to enumerate subgraphs of the bidirectional relationship graph. It also uses a depth-first traversal to generate a set of connected subgraphs and uses the set of style elements corresponding to each connected subgraph as a candidate style combination scheme. The third unit is used to calculate the cosine similarity between the context-enhanced feature vectors corresponding to different style elements within each candidate style combination scheme to obtain the element similarity matrix, calculate the matrix norm of the element similarity matrix and determine the internal consistency score of the scheme, extract the topological characteristic parameters of the candidate style combination scheme in the bidirectional relationship graph and calculate the scheme structure stability score. The fourth unit is used to calculate the combination coordination score of each candidate style combination scheme based on the internal consistency score and the structural stability score of the scheme, and to select the target style combination scheme based on the combination coordination score.

5. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 3.

Images