A clothing style perception evaluation method based on multi-modal feature fusion

Abstract

The application discloses a kind of clothing style perception evaluation methods based on multi-modal feature fusion, specifically related to computer vision and multi-modal data fusion technical field, including the acquisition of the same clothing object's show record, extract visual vector with area position number from clothing image using deep learning, encode text into text semantic vector, write in the area position number pointed by scene label after adjusting direction by age layer, rewrite numerical value according to comment emotion symbol;According to the order of display, stack visual vector and embed text semantic vector, form feature tensor and make tree tensor network decomposition, with display order as tree root, area position number as leaf, enumerate sub-tensor rank and make singular value decomposition;By the area alignment fusion of the multiple show records of the same clothing object, tree tensor network decomposition, reconstruction residual flow and distortion riemannian metric learning, clothing ontology features and matching incremental features are used for style evaluation, and interference features formed by display conditions are excluded.

Description

Technical Field

[0001] This invention relates to the field of computer vision and multimodal data fusion technology, and more specifically, to a method for evaluating clothing style perception based on multimodal feature fusion. Background Technology

[0002] Existing methods for evaluating clothing style generally extract color, cut, material, and matching relationships from clothing images through edge computing, extract semantic information from clothing text, user age groups, scene tags, and comments, and use features from different sources together to evaluate fashionability, age suitability, scene matching, and aesthetic acceptance. In e-commerce and social platforms, the same clothing is often displayed repeatedly with different models, outfits, backgrounds and filters. When edge computing nodes receive display data, they must retain the effective information formed by the outfit relationship, while eliminating the influence of model body shape, shooting environment and filter color deviation. They cannot directly use background removal, accessory removal or color normalization processing. Existing edge computing methods often use a single published image and associated text as a set of inputs, which easily leads to clothing features, matching features, and display features being written into the fusion result. Specifically, the evaluation result changes when the same clothing is changed by changing the model, background, or filter, and the source of the change cannot be determined. The reason for this problem is that there is a lack of feature preservation and feature change comparison based on the same clothing between different display records. The technical problem this application aims to solve is: while preserving the true matching information, using edge computing to process multiple display data of the same garment, separating the interference caused by display conditions, and obtaining a style evaluation result supported by the garment design features and matching relationships. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a clothing style perception evaluation method based on multimodal feature fusion. By performing region alignment fusion, tree tensor network decomposition, reconstruction residual splitting, and distorted Riemann metric learning on multiple display records of the same clothing object, the clothing ontological features and matching incremental features are used for style evaluation, and interference features formed by display conditions are eliminated, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for evaluating clothing style perception based on multimodal feature fusion, comprising: S1. Obtain the display records of the same clothing object, use deep learning to extract visual vectors with region identifiers from the clothing images, encode the text into text semantic vectors, adjust the direction by age group and write it into the region identifier pointed to by the scene label, and rewrite the value according to the comment sentiment symbol. S2. Stack visual vectors according to the display order and embed text semantic vectors to form feature tensors and perform tree tensor network decomposition. With the display order as the root and the region position as the leaf, enumerate the sub-tensor ranks and perform singular value decomposition. Determine the shrinking rank according to the sum of the number of identical signs in the reconstruction residual and the number of region position, and generate a tree tensor network. S3. Set each parent node to zero and reconstruct the display record through the neural decoder. Write the clothing body tree when the residual sign and absolute value are the same as the non-zero residual number. Otherwise, write the matching increment tree when the parent node area is connected to the boundary of the clothing body tree node area and the occlusion order remains unchanged. Write the remaining parent nodes into the display interference tree. S4. Perform Twisted Riemann metric learning on the tree nodes. After mapping by the neural encoder, multiply the Jacobi matrix of the neural decoder by the transpose matrix to generate the pulled-back Riemann metric. Based on the reconstruction residuals of the clothing ontology tree and the display interference tree, shrink the ontology mapping position and expand the interference mapping position. Generate the twisted Riemann distance through path bisection and energy summation. S5. Based on the distorted Riemann distance, the text semantic vector is written into the first tree node in ascending order of distance. The text semantic vector in the display interference tree is deleted and the interference mapping position is expanded according to the number of deletions. The writing and expansion are repeated according to the tree level. The clothing ontology tree and the matching increment tree are shrunk to form a style fusion vector. The clothing style perception evaluation result is output through a deep classification neural network.

[0005] In a preferred embodiment, S1 includes: S11. For a single display record with the same clothing object encoding, the deep convolutional network reads the clothing image and generates a pixel feature map. The human body key point branch outputs the key point coordinates according to the human skeleton order, and the clothing segmentation branch outputs the clothing contour mask. The key point coordinates are projected horizontally onto the boundary of the clothing contour mask. The region number is generated in ascending order according to the vertical coordinate of the projection hit point. Then, the pixel features covered by the region number are summed dimension by dimension and divided by the number of covered pixels to output a visual vector with region number. S12. For related text in a single display record, the transform coding neural network outputs word vectors in word order, generates a segment semantic vector by the average of word vectors between sentence-end markers, and the age embedding network expands the age layer encoding into an age vector. The segment semantic vector and the age vector are multiplied bit by bit and then summed to obtain the age inner product. When the age inner product is negative, the segment semantic vector is inverted bit by bit. When the age inner product is zero or positive, the segment semantic vector is preserved, and the age-adjusted semantic vector is output. S13. The scene embedding network expands the scene label into a scene vector. The scene vector is multiplied bit by bit by each visual vector and then summed. The sum is sorted in descending order and in ascending order by region number when the sums are the same. The bidirectional long short-term memory network reads the comment segment and generates a sentiment difference through a linear layer. When the sentiment difference is negative, it is taken as negative one, and when the sentiment difference is zero or positive, it is taken as one. The value is multiplied into the age-adjusted semantic vector and then written into the region number. This generates a display semantic vector that is aligned and fused with the visual vector according to the region number.

[0006] In a preferred embodiment, S2 includes: S21. For each region number under the display order, read the visual vector and text semantic vector corresponding to the region number. When the dimensions are inconsistent, pad the tail of the short-dimensional vector with zeros until the dimensions are consistent. Then add them one by one and record the region number to generate the region fusion vector. S22. Using the display order as the first dimension, the region position number as the second dimension, and the vector position number as the third dimension, write the region fusion vector into the corresponding tensor cell. Write zero values ​​into the tensor cells that have not obtained the region fusion vector and mark them as missing, thus generating the feature tensor.

[0007] In a preferred embodiment, S2 further includes: S23. Perform tree tensor network decomposition on the feature tensor to connect the root of the tree in the display order and connect the leaves with the region position number. Merge the sub-tensors according to the adjacent region position numbers. Expand each sub-tensor into a matrix along the display order. When the number of rows in the matrix does not exceed the number of columns, use the number of rows as the rank enumeration endpoint. When the number of rows in the matrix exceeds the number of columns, use the number of columns as the rank enumeration endpoint. Perform singular value decomposition on each integer rank between the first and the rank enumeration endpoint to generate candidate parent node kernels. S24. Reconstruct the corresponding sub-tensor using candidate parent node kernels. Subtract the original tensor value from the reconstructed value to form the reconstruction residual. Take 1 when the reconstruction residual is positive, take -1 when it is negative, and take 0 when it is zero. Sort the candidate parent node kernels in descending order according to the sum of the number of residuals with the same sign in the same region and the number of regions with the same sign after reconstruction. When they are in the same position, take the first candidate parent node kernel in ascending order of integer rank to shrink the sub-tensor. Shrink the sub-tensor layer by layer from the leaves to the root and output the tree tensor network.

[0008] In a preferred embodiment, S3 includes: S31. For the parent node in the tree tensor network, set the parent node kernel to zero according to the parent node order and reconstruct and display the record through the neural decoder. Read the difference of the fusion vector under the same region position before and after setting it to zero. Summate the fusion vector difference dimensionally to obtain the region residual. Write 1 if the region residual is positive, write negative 1 if it is negative, and write zero if it is zero. Generate the residual sign bit and record the region position and occlusion order after reconstruction. S32. Based on the residual sign bit of the same parent node in all displayed records, count the number of non-zero region residuals, the absolute number of the cumulative residual sign bit value, and the change in region position number. Generate the ontology attribution value by subtracting the absolute number of the cumulative residual sign bit value from the number of non-zero region residuals and adding the change in region position number.

[0009] In a preferred embodiment, S3 further includes: S33. Generate the initial node of the clothing ontology tree from the parent node with the ontology attribution cost value of zero. For the parent node with the ontology attribution cost value of non-zero, calculate the number of the same coordinates between its region boundary coordinates and the region boundary coordinates of the initial node of the clothing ontology tree. Determine the successor node in descending order of the number of the same coordinates. Form the boundary missing position with the number of the same coordinates of zero and form the occlusion offset value with the absolute value of the occlusion order difference. Add the boundary missing position and the occlusion offset value to generate the matching attribution cost value. S34. Take the sum of the attribution costs of all parent nodes as the first target bit and the number of matching attributions without successor nodes as the second target bit. Determine the attribution of parent nodes by the first position in ascending order of the target words. Among them, the parent nodes with the attribution cost of the main body are written into the clothing main body tree, the parent nodes with the attribution cost of the matching are written into the matching incremental tree, and the remaining parent nodes are written into the display interference tree. The first round takes the attribution result of the previous round as all zeros as input. When the attribution result is XORed with the attribution result of the previous round node by node and the sum is zero, the tree attribution result is output. Otherwise, the number of the same coordinates and the occlusion order difference are recalculated to the tree level based on the attribution result.

[0010] In a preferred embodiment, S4 includes: S41. Based on the tree attribution results, read the node vector, region number, source position and reconstruction residual of each tree node. The neural encoder maps the node vector to the latent coordinates. The neural decoder generates the Jacobian matrix at the latent coordinates and multiplies the Jacobian matrix by the transpose matrix and adds one to the diagonal to form the basic metric of the node. S42. For the same region reference number, sum the absolute values ​​of the reconstruction residuals where the node's attribute word is the ontology value to obtain the ontology residual sum, and sum the absolute values ​​of the reconstruction residuals where the node's attribute word is the interference value to obtain the interference residual sum. Add one to the ontology residual sum and divide by the ontology residual sum plus the interference residual sum plus two to form a contraction factor. Add two to the ontology residual sum plus the interference residual sum and divide by the interference residual sum plus one to form an expansion factor. Use the contraction factor to rewrite the basic metric of the ontology value node, and use the expansion factor to rewrite the basic metric of the interference value node.

[0011] In a preferred embodiment, S4 further includes: S43. For visual source nodes and text source nodes in the same region, calculate the ontology path energy from the text source node to the ontology value node and the interference path energy from the text source node to the interference value node under the rewritten node base metric. Subtract the interference path energy from the ontology path energy to form the path energy difference. When the path energy difference is negative, increment the ontology support count by one. When the path energy difference is zero, increment both the ontology support count and the interference support count by one. When the path energy difference is positive, increment the interference support count by one. Then, divide the incremented ontology support count by the sum of the ontology support count and the interference support count by two to generate the multi-source fusion confidence word. S44. Multiply the multi-source fusion confidence word into the basic metric of the ontology value node, multiply the result of subtracting the multi-source fusion confidence word from one and then adding one into the basic metric of the interference value node, perform path bisection on the adjacent potential coordinates according to the tree edge connection order, take the average of the two potential coordinates at the midpoint, and obtain the path energy of this round by the product of the adjacent coordinate difference vector, the node basic metric and the transpose of the adjacent coordinate difference vector. S45. Record the path energy in the first round. Starting from the second round, stop the distorted Riemann metric learning when the fixed value of the path energy in the current round is the same as the fixed value of the path energy in the previous round, or when the number of path bisection rounds is equal to the number of tree tensor network layers. The square root of the path energy in the current round is used as the distorted Riemann distance output.

[0012] In a preferred embodiment, S5 includes: S51. Based on the distorted Riemann distance table, read the distance fixed point value in groups according to the text number, rearrange the candidate tree nodes with the same distance fixed point value in ascending order according to the node path code, take the path code of the first node in ascending order, the corresponding region number and the tree node belonging word to concatenate the classification word, and generate the current round of classification table. S52. Read the tree node's belonging word according to the current round's inclusion table. When the belonging word is an interference value, write the corresponding text number into the deletion table and increment the deletion count in the same area. When the belonging word is a non-interference value, sum the text semantic vector and the tree node vector bit by bit and divide by two to generate the node fusion vector.

[0013] In a preferred embodiment, S5 further includes: S53. Update the interference mapping coordinates based on the deletion count. First, calculate the average value of the body mapping coordinates in the same region to form the body center coordinates. Then, subtract the body center coordinates from the interference mapping coordinates to obtain the offset vector. Multiply the offset vector by the deletion count plus one and add it back to the body center coordinates. Recalculate the distorted Riemann distance and return to S51. Stop when each of the input words is equal to the input word of the previous round. If they are not the same, stop when the number of rounds reaches the tree level. Output the iterative input table. S54. Based on the iterative induction table, retain the node fusion vector of the final round, perform tree tensor contraction from the leaves to the root, multiply the child node fusion vector under the same parent node bit by the parent node kernel coefficient and sum them. When the absolute value of the coefficient is non-zero, divide it by the sum of the absolute values ​​of the coefficients. When the sum of the absolute values ​​of the coefficients is zero, write the zero vector to obtain the root node style fusion vector. Then, perform classification-level fusion by the deep classification neural network and output the clothing style perception evaluation result.

[0014] The technical effects and advantages of this invention are as follows: This solution uses tree tensor decomposition and residual splitting to separate the ontology, collocation, and interference, thereby relatively reducing the impact of background, filter, and model differences on the evaluation. By aligning and fusing images and text with region identifiers, semantic meaning is written into the corresponding clothing region, reducing the occurrence of text without region pointers entering style fusion. After the parent node is set to zero, the display record is reconstructed by the edge computing node. The ontology is determined by residual consistency and region preservation, which reduces the sensitivity of ontology evaluation to display changes. Boundary continuity and occlusion order constraints, combined with incremental attribution, allow the actual dressing relationship to participate in the evaluation, relatively reducing the information loss caused by removing accessories. The residual-driven distortion Riemann metric shrinks the ontology and expands the interference, causing text to be incorporated into the effective region, and relatively suppressing the mixing of comment noise into the edge computing fusion result. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention. Detailed Implementation

[0016] 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.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides a method for evaluating clothing style perception based on multimodal feature fusion, comprising: S1. Obtain the display records of the same clothing object, use deep learning to extract visual vectors with region identifiers from the clothing images, encode the text into text semantic vectors, adjust the direction by age group and write it into the region identifier pointed to by the scene label, and rewrite the value according to the comment sentiment symbol. In this implementation, S1 performs image region extraction and text region writing for display records with the same clothing object code. First, the clothing image is converted into a visual vector with a region identifier, and then the associated text is converted into a display semantic vector that is jointly modified by age group, scene tag, and comment sentiment. This allows S2 to read the visual vector and display semantic vector according to the display order and region identifier. The clothing object code is read first by the product style number. If the product style number is empty, the same style release identifier is read. If the same style release identifier is empty, the same style aggregation result is read. When the same display record contains multiple clothing object codes, only the records where the clothing object code to be evaluated matches are read. Display records where no clothing object code is read are written as removal records. This implementation process includes the following steps: In S11, the clothing image is transformed into a visual vector arranged by clothing regions. The region identifier is generated by the projection relationship between human keypoints and clothing contours. The deep convolutional network scales the clothing image according to the training input size and converts the pixel coordinates. The backbone network outputs pixel feature maps. The human keypoint branch outputs keypoint coordinates according to the human skeleton order. The clothing segmentation branch outputs clothing contour masks. The connected component with the first-to-last area in descending order and that is encoded by the clothing object is written into the target contour. The connected component that has a boundary contact with the target contour is written into the matching candidate contour. The remaining connected components do not participate in the generation of region identifiers. Project the key point coordinates onto the target contour boundary using the horizontal scan line. When the projection hits the left or right boundary, take the hit point with the closer horizontal coordinate distance. When the projection misses the target contour boundary, take the boundary point with the first Euclidean distance in ascending order. Generate the region number according to the vertical coordinate of the hit point in ascending order. If the vertical coordinates are the same, arrange them according to the horizontal coordinate in ascending order. The target contour pixels between the ordinates of adjacent hit points are written into the corresponding region identifiers. The pixels above the first hit point are written into the first region identifier, and the pixels below the last hit point are written into the last region identifier. The pixel features covered by each region identifier are summed dimension by dimension and divided by the number of covered pixels to obtain a visual vector, which is then written into the visual vector table. When keypoint coordinates are missing, the average coordinates of adjacent skeleton keypoints are read and replaced. When the target contour is missing, a zero visual vector is output and a visual absence marker is written. When the number of covered pixels is zero, a zero visual vector is output and a region absence marker is written. In S12, the associated text is transformed into an age-adjusted semantic vector constrained by the age layer direction. The associated text reads the title field and outfit description field from the display record and concatenates them according to the field order. The transform coding neural network outputs word vectors according to word order. The punctuation segmentation result forms the sentence end mark. When the punctuation segmentation result is empty, the last word of the associated text is used as the sentence end mark. The word vectors between adjacent sentence end marks are summed dimension by dimension and divided by the number of words to generate a segment semantic vector and write it into the text number. The age embedding network expands the age layer encoding into an age vector with the same dimension as the word segment semantic vector. The word segment semantic vector and the age vector are multiplied bit by bit and then summed to obtain the age inner product. When the age inner product is negative, the word segment semantic vector is inverted bit by bit. When the age inner product is zero or positive, the word segment semantic vector is preserved, and the age-adjusted semantic vector is obtained and written into the age-adjusted semantic vector table according to the text number. When the associated text is empty, a zero word segment semantic vector is output and a text missing marker is written. When the age layer encoding is empty, the age vector is written with all 1s, so that the age-adjusted semantic vector follows the value of the word segment semantic vector. In S13, the age-adjusted semantic vector is written into the clothing area selected by the scene label, and the numerical direction is changed by the comment sentiment to form the display semantic vector; the scene embedding network expands the scene label into a scene vector, multiplies the scene vector with each visual vector in the visual vector table bit by bit and sums them to obtain the region sum value, and obtains the writing region number in descending order of the region sum value. If the region sum values ​​are the same, they are obtained in ascending order of the region number. The comment content is segmented into comment segments based on the sentence-end marker. After the bidirectional long short-term memory network reads the comment segments, it outputs positive and negative sentiment values ​​through a linear layer. The positive sentiment value is subtracted from the negative sentiment value to generate a sentiment difference. When the sentiment difference is negative, it is taken as negative one, and when the sentiment difference is zero or positive, it is taken as one. The value is multiplied by the age to adjust the semantic vector and then written into the writing area number obtained from the scene label to generate a display semantic vector carrying the text number, area number and vector value. When the scene tag is empty, the scene vector is written to zero and the first one is obtained and written to the region position number in ascending order. When the comment content is empty, the emotion value is one. When there are multiple display semantic vectors under the same region position number, they are written to the display semantic vector table according to the text number and the original vector is not overwritten. Through the above processing, the clothing image forms a visual vector organized by region position number, and the associated text forms a display semantic vector organized by text number and region position number. The visual vector table and the display semantic vector table are used together by S2 to construct the feature tensor. The visual missing marker, region missing marker and text missing marker are used in S2 for zero value writing and count exclusion. In practical applications: There are multiple display records for the same dress with the same product style number. The deep convolutional network divides the clothing image into the neck, waist and hem areas according to human body key points. The phrase "slimming for commuting" in the outfit description is generated into a segment semantic vector by a transform coding neural network. The age layer encoding changes the direction of the segment semantic vector. The scene label writes the changed vector into the waist area. The emotional direction in the comment content is rewritten and written into the numerical value, and finally the display semantic vector read by S2 is formed.

[0018] S2. Stack visual vectors according to the display order and embed text semantic vectors to form feature tensors and perform tree tensor network decomposition. With the display order as the root and the region position as the leaf, enumerate the sub-tensor ranks and perform singular value decomposition. Determine the shrinking rank according to the sum of the number of identical signs in the reconstruction residual and the number of region position, and generate a tree tensor network. In this embodiment, the image-side values ​​and text-side values ​​of the same region number under the same display order are synthesized into a region fusion vector. Then, the display order, region number, and vector number are written into a feature tensor. The feature tensor is compressed into a hierarchical structure with parent node kernels through tree tensor network decomposition, enabling subsequent step S3 to set parent nodes to zero one by one and track the reconstruction residuals. The display order is generated in ascending order of display record publication time; if the publication times are the same, it is generated in ascending order of display record number. The region number uses the result generated in S1, and the vector number is generated in order of vector dimension. This implementation process includes the following steps: In S21, the visual vector and display semantic vector under the same region position number are synthesized into a region fusion vector, which is used as the writing unit of the feature tensor. The visual vector table is read according to the display order, and then the display semantic vector table is read according to the same display order and the same region position number. If there are multiple display semantic vectors under the same region position number, they are summed dimension by dimension in ascending order of text number and divided by the number of display semantic vectors to obtain the average regional semantic vector. Compare the dimension of the visual vector and the dimension of the region semantic mean vector. Pad the shorter vector with zeros at the end to the dimension of the longer vector. Then add the two zero-padding vectors bit by bit to generate a region fusion vector carrying the display order and region position number. If the display semantic vector is not read, use the zero vector of the same dimension as the region semantic mean vector and write a text missing mark. If the visual vector is not read, use the zero vector of the same dimension as the visual vector and write a visual missing mark. The visual missing mark and the text missing mark enter S22 along with the region fusion vector. In S22, the region fusion vector is written into a three-dimensional tensor grid so that the tree tensor network decomposition can simultaneously process multi-source features according to the display record dimension and the clothing region dimension. The display order is used as the first dimension index, the region position number is used as the second dimension index, and the vector position number is used as the third dimension index. The vector position number of the region fusion vector is written into the corresponding tensor grid bit by bit to obtain the feature tensor. For regions lacking region fusion vectors under the same display priority, write a zero value and a tensor missing marker to the corresponding tensor cell. The tensor missing marker is generated by bitwise ORing the visual missing marker and the text missing marker. When a written value already exists in the same tensor cell, retain the written value with the earlier display priority and write the later written value to the duplicate tensor cell record. The tensor missing marker is not counted when the number of reconstructed residuals in the same direction and the number of region missing markers are kept in S24. In S23, the feature tensor is decomposed into a tree tensor network that shrinks from the region leaf nodes to the display root node. The parent node kernel is generated by the singular value decomposition result corresponding to the candidate integer rank. First, adjacent region position numbers are merged into sub-tensors in ascending order of region position numbers. When the number of region position numbers is odd, the last region position number is merged with the zero sub-tensor. All tensor cells of the zero sub-tensor are written with zero values ​​and a merge missing mark is written. Each sub-tensor is expanded into a matrix along the display order. The matrix rows correspond to the display order, and the matrix columns correspond to the concatenation of the region index and vector index within the merged region. The numerical items in the matrix row number and matrix column number serve as the end point of rank enumeration. When the row number or column number is zero, the current sub-tensor is skipped and written to the empty parent node record. Perform singular value decomposition on each integer rank between the end points of the rank enumeration and the rank, take the truncated left singular vector, truncated singular value and truncated right singular vector under the corresponding integer rank and multiply them to generate candidate parent node kernels, and write the candidate parent node kernels into the parent node kernel coefficient table. In S24, the candidate parent node kernel reverse-selects the shrinking rank through the reconstruction result, so that the tree tensor network retains the characteristic structure of stable regional positions in the same display record; the corresponding child tensor is reconstructed using the candidate parent node kernel, and the reconstruction value is subtracted from the original tensor value to form the reconstruction residual. When the reconstruction residual is positive, it is written as 1; when it is negative, it is written as negative 1; when it is zero, it is written as zero. The number of records with the same statistical sign of the reconstructed residuals without tensor missing markers under the same original region position number is used to obtain the number of residuals with the same direction; each reconstructed vector is multiplied bit by bit with the original region fusion vectors under the same display order and then summed. The reconstructed region position number is obtained in descending order of the sum value. If the sum values ​​are the same, the reconstructed region position number is obtained in ascending order of the region position number. If the reconstructed region position number is the same as the original region position number, the region position number is incremented by one. The candidate sorting value is obtained by adding the number of residual signs in the same direction and the number of region positions. The candidate sorting values ​​are arranged in descending order. When the scores are the same, the first candidate parent node kernel is obtained in ascending order of integer rank as the parent node kernel of the current sub-tensor. The process of merging, expanding, enumerating, rebuilding and shrinking is repeated layer by layer from the leaves to the root of the tree to output the tree tensor network and the parent node kernel coefficient table. Through the above processing, the visual vector and the display semantic vector form a region fusion vector on the region position number. The region fusion vector is written into the feature tensor according to the display order, region position number and vector position number. The feature tensor is decomposed into parent node kernel and tree node path through the tree tensor network. The parent node kernel coefficient table is used by S3 to execute the parent node to zero. The tensor missing mark is used by S3 and S4 to eliminate the residual disturbance caused by missing data. In practical applications: Multiple display records of the same dress are generated according to the release time to form a display order. The neckline, waist and hem areas are written with area numbers respectively. The visual vector and the display semantic vector corresponding to "commuting slimming" are added in the waist area to form an area fusion vector. The area fusion vectors of multiple display orders are written into the feature tensor. The tree tensor network decomposes and generates parent node kernels layer by layer between the neckline and waist, and between the waist and hem. Subsequently, S3 determines whether the corresponding feature belongs to the clothing body, matching increment or display interference by setting the parent node kernel to zero.

[0019] S3. Set each parent node to zero and reconstruct the display record through the neural decoder. Write the clothing body tree when the residual sign and absolute value are the same as the non-zero residual number. Otherwise, write the matching increment tree when the parent node area is connected to the boundary of the clothing body tree node area and the occlusion order remains unchanged. Write the remaining parent nodes into the display interference tree. In this implementation, by setting each parent node core to zero and reconstructing the display record, the influence of the parent node on the region fusion vector, region position number, and occlusion order in the same display record is calculated. Then, the parent nodes are assigned to the clothing ontology tree, the matching incremental tree, or the display interference tree. The parent node order is generated in ascending order according to the node path encoding. In the tree node's attribution word, the value for unattributed nodes is written as zero, the value for ontology nodes is written as one, the value for matching nodes is written as two, and the value for interference nodes is written as three. This implementation process includes the following steps: In S31, the parent node kernels are set to zero one by one to obtain the influence of a single parent node on the reconstruction result of the display record; the input is a tree tensor network, a parent node kernel coefficient table, a feature tensor and a tensor missing marker. During processing, the node path code is read in order of parent node, and all kernel coefficients under the same node path code in the parent node kernel coefficient table are rewritten to zero, while the kernel coefficients of other parent nodes are kept at their original values. The neural decoder outputs the reconstruction record package based on the rewritten parent node kernel coefficient table. The reconstruction record package includes the reconstruction fusion vector, the reconstructed region ID, and the reconstruction occlusion order for each display order. The reconstructed region ID is obtained by multiplying the reconstruction fusion vector with each original region fusion vector under the same display order and summing the results. The region ID corresponding to the first position of the summation value in descending order is written as the reconstructed region ID. If the summation values ​​are the same, the region IDs are taken in ascending order. The occlusion order is read from the hierarchical channel of the clothing segmentation branch. If the hierarchical channel is not hit, it is generated in descending order according to the overlapping area of ​​the region boundary. If the overlapping areas are the same, it is generated in ascending order according to the region ID. For the same display order and the same region position number, read the original region fusion vector before zeroing and the reconstructed fusion vector after zeroing. Subtract the latter from the former to form the fusion vector difference. Summate the fusion vector difference dimension by dimension to obtain the region residual. Write 1 if the region residual is positive, write -1 if it is negative, and write 0 if it is zero. Generate the residual sign bit and write it into the residual table of the parent node. When the region position number has a tensor missing mark, write 0 into the region residual and write the residual missing mark into the residual table of the parent node. In S32, the residual sign bits of the same parent node in all display records are summarized into the ontology attribution value, which is used to indicate whether the parent node maintains the same direction of influence in the same display record; the input is the parent node residual table, the reconstructed region position number and the region position number before zeroing. During processing, the region residuals without residual missing mark are read in groups according to the parent node path code. When the region residual is not zero, the number of non-zero region residuals is incremented by one, and the corresponding residual sign bits are accumulated to form the residual sign bit accumulation value. When the reconstructed region number under the same display order is different from the region number before zeroing, the change in region number is incremented by one; when they are the same, the change in region number remains unchanged. Subtract the absolute number of the cumulative value of the residual sign bit from the number of non-zero region residuals, and then add the change in region number to obtain the ontology attribution cost and write it into the parent node attribution cost table. When the same parent node has a residual missing mark in all displayed records, the number of non-zero region residuals, the absolute number of the cumulative value of the residual sign bit, and the change of region position number are all written to zero. The parent node's attribution cost table is synchronously written with the full missing mark for S34 to read during attribution sorting. In S33, boundary continuation and occlusion order are used to identify parent nodes that have a matching relationship with the clothing body region. The inputs are the parent node attribution cost table, the region boundary coordinate table, the reconstructed occlusion order, and the node path code. During processing, the parent nodes with a body attribution cost value of zero and without a full missing marker are read first to generate the initial node of the clothing body tree. When the parent node with a body attribution cost value of zero is empty, the parent node with the first position in ascending order of body attribution cost value and the region position number is kept non-zero is read as a temporary body node. The temporary body node is only used for continuation calculation. The region boundary coordinates are formed by the integer pixel coordinates of the mask boundary corresponding to the region number. When the parent node covers multiple region numbers, the union of multiple region boundary coordinate sets is taken. For parent nodes whose ontology ownership value is not zero, the intersection of the parent node's region boundary coordinate set and the initial node's region boundary coordinate set of the clothing ontology tree is calculated. The number of intersection coordinates is written as the number of identical coordinates. The successor candidate nodes are read in descending order of the number of identical coordinates. When the number of identical coordinates is the same, the successor candidate nodes are read in ascending order of the node path code. When the number of identical coordinates is not zero, the successor candidate node is written into the successor node field. When the number of identical coordinates is zero, the successor node field is written empty and the missing boundary bit is written as 1. The occlusion offset value is obtained by subtracting the absolute value of the occlusion order of the successor node from the occlusion order of the parent node. When the successor node field is empty, the occlusion offset value is written as 1. The missing boundary position is added to the occlusion offset value to generate the matching attribution cost value and written to the parent node attribution cost table. In S34, the parent node attribution is determined by sorting the target words and iterative recalculation, and the tree attribution result is output when the attribution result no longer changes. The input is the parent node attribution cost table, the successor node field, the tree level and the previous round attribution result. The first round of the previous round attribution result is generated based on the values ​​of all parent nodes that are not attributed. During processing, the ontology attribution cost and the combination attribution cost are read for each parent node. When the ontology attribution cost is zero and there is no full absence mark, it is written into the clothing ontology tree. When the combination attribution cost is zero and the successor node field is not empty, it is written into the combination incremental tree. The remaining parent nodes are written into the display interference tree. After each round of attribution, the write value in the ontology attribution cost value and the combination attribution cost value are accumulated according to the parent node order to form the first target bit. The number of parent nodes whose successor node field is empty but have generated a combination attribution cost value is counted to form the second target bit. The first target bit comes first and the second target bit comes last and are concatenated to form the target word. When multiple attribution candidates are generated in the same round, the first attribution candidate is selected in ascending order of the target word. When the target words are the same, the attribution candidates are selected in ascending order of the parent node path code. The current round's assignment result is XORed with the previous round's assignment result node by node and summed. When the sum is zero, the tree assignment result is output. When the sum is not zero, the number of the same coordinates and the occlusion order difference in S33 are recalculated based on the current round's assignment result, and the next round is entered. When the round reaches the tree level, the tree assignment result corresponding to the first ascending word of the target word is output. Through the above processing, the parent node in the tree tensor network obtains the tree node attribution word formed by residual direction, region migration, boundary continuation and occlusion order. The clothing ontology tree, matching incremental tree and display interference tree are respectively written into the path code of the corresponding parent node. The tree attribution result is used by S4 to read the node vector, region number, source bit and reconstruction residual. In practical applications: In the display records of the same dress, after the parent node corresponding to the waist pattern is set to zero, a residual in the same direction is generated in multiple display orders and the region position remains unchanged. The parent node is written into the garment body tree. After the parent node of the belt is set to zero, the reconstruction fusion vector is changed only in some display orders. However, the boundary of the belt region is connected to the waist region and maintains the occlusion order. The parent node is written into the matching increment tree. After the parent node of the background setting is set to zero, the region position migrates and there is no successor node. The parent node is written into the display interference tree.

[0020] S4. Perform Twisted Riemann metric learning on the tree nodes. After mapping by the neural encoder, multiply the Jacobi matrix of the neural decoder by the transpose matrix to generate the pulled-back Riemann metric. Based on the reconstruction residuals of the clothing ontology tree and the display interference tree, shrink the ontology mapping position and expand the interference mapping position. Generate the twisted Riemann distance through path bisection and energy summation. In this embodiment, tree nodes in the clothing ontology tree, the matching increment tree, and the display interference tree are mapped to the latent space. The reconstruction residual is used to change the basic metric of the nodes in the latent space, causing the text source node to converge towards the location of the clothing ontology tree or the matching increment tree, and causing the location of the display interference tree to move away from the text source node. The tree node attribution word uses the ontology value, matching value, and interference value from S3. The node path encoding uses the root-to-tree node branch order generated in S2. The source bit is generated by the case of the visual vector and the display semantic vector participating in the region fusion vector in S21. This implementation process includes the following steps: In S41, tree nodes are converted into latent coordinates with node basic metrics for subsequent residual shrinking, residual expansion, and path energy calculations. The inputs are tree attribution results, node vectors, region identifiers, source identifiers, reconstruction residuals, and node path codes. During processing, the node vector of each tree node is read, the neural encoder maps the node vectors to latent coordinates, and the neural decoder takes the latent coordinates as input and outputs the node reconstruction vector. The rows of the Jacobi matrix correspond to the latent coordinate dimensions, and the columns of the Jacobi matrix correspond to the node reconstruction vector dimensions. After multiplying the Jacobi matrix by the transpose of the Jacobi matrix, each of the main diagonal elements of the product matrix is ​​incremented by one to generate the node basic metrics and write them into the node metric table. When the visual source position in the source position is one, a visual source node is generated. When the text source position in the source position is one, a text source node is generated. When both the visual source position and the text source position are one, they are copied into visual source copy and text source copy according to the same node path encoding. When the node vector is empty, a zero node vector is written and a node missing mark is written. The tree node corresponding to the node missing mark does not enter the path energy calculation of S43. In S42, the reconstruction residuals under the same region reference number are converted into a shrinkage factor and an expansion factor, which shorten the path length of the location of the ontology value node and increase the path length of the location of the interference value node. The inputs are the node metric table, the tree node attribution word, the region reference number, and the reconstruction residuals. During processing, the same region reference number is grouped, and the absolute values ​​of the reconstruction residuals with the tree node attribution word being the ontology value are summed to obtain the ontology residual sum. The absolute values ​​of the reconstruction residuals with the tree node attribution word being the interference value are summed to obtain the interference residual sum. The value nodes do not participate in the summation of the two terms. The contraction factor is generated by dividing the sum of the ontology residuals plus one by the sum of the ontology residuals, the sum of the interference residuals and two. The expansion factor is generated by dividing the sum of the ontology residuals, the sum of the interference residuals and two by the sum of the interference residuals plus one. The contraction factor is multiplied into the basic metric of the ontology value node, and the expansion factor is multiplied into the basic metric of the interference value node. The results of S41 are used in combination with the basic metric of the value node. Both the sum of the original residual and the sum of the interference residual are recorded according to a fixed number of points. The addition of one and the addition of two are generated by integer constants. When there is no original value node under the region reference number, the sum of the original residual is written as zero. When there is no interference value node under the region reference number, the sum of the interference residual is written as zero. In S43, the path energy of the text source node relative to the ontology value node and the interference value node is converted into a multi-source fusion confidence word, which is used in S44 to rewrite the node's basic metric again. The input is the rewritten node basic metric, potential coordinates, region number and source number. During processing, the text source node, visual source node, ontology value node and interference value node are read within the same region number. The ontology path energy from the text source node to the ontology value node is calculated, and the interference path energy from the text source node to the interference value node is calculated. The path energy is obtained by multiplying the potential coordinate difference vectors at both ends, the mean matrix of the basic metrics of the nodes at both ends, and the transpose matrix of the difference vectors in sequence. When there are multiple ontology value nodes in the same region, the first one is selected in ascending order of ontology path energy. When there are multiple interference value nodes in the same region, the first one is selected in ascending order of interference path energy. The path energy difference is formed by subtracting the interference path energy from the ontology path energy. When the path energy difference is negative, the ontology support count is incremented by one. When the path energy difference is zero, both the ontology support count and the interference support count are incremented by one. When the path energy difference is positive, the interference support count is incremented by one. The multi-source fusion confidence word is generated by dividing the incremented ontology support count by the sum of the ontology support count, the interference support count, and two. When the same region bit is missing a text source node, ontology value node, or interference value node, the corresponding region bit does not update the support count and writes a path missing marker. In S44, the multi-source fusion confidence word is written back to the node base metric, and the path energy of this round is formed by tree edge path binary search. The input is the multi-source fusion confidence word, the node base metric, the potential coordinates and the tree edge connection order. During processing, the multi-source fusion confidence word is multiplied into the ontology value node base metric, and the result of subtracting the multi-source fusion confidence word from one and adding one is multiplied into the interference value node base metric. The current value is maintained by combining the value node base metric. Read the adjacent potential coordinates according to the tree edge connection order, perform path bisection for each tree edge, and take the midpoint coordinate by adding the potential coordinates of the two ends one by one and then dividing by two, and insert the midpoint coordinate between the potential coordinates of the two ends. For each pair of adjacent coordinates in the path, first calculate the adjacent coordinate difference vector, then read the mean matrix of the basic metric of the nodes at both ends of the adjacent coordinates, and multiply the adjacent coordinate difference vector, the mean matrix, and the transpose matrix of the adjacent coordinate difference vector in turn to obtain the energy of the adjacent path segment. Summing up the energy of all adjacent path segments gives the energy of the current round of the path. When the tree edge connection order is empty, write zero to the energy of the current round of the path and write the path empty marker. In S45, the path energy is compared in each round to form a stopping result, and the path energy of the stopping round is converted into a twisted Riemann distance and output to S5. The inputs are the path energy of the current round, the path energy of the previous round, the number of tree tensor network layers, and the path binary search round. In the first round, only the path energy of the current round is recorded. From the second round onwards, the path energy of the current round is amplified and rounded according to the decimal places of the input vector to obtain the fixed-point value of the path energy of the current round. The path energy of the previous round is amplified and rounded according to the same decimal places to obtain the fixed-point value of the path energy of the previous round. Stop when the path energy fixed point value of the current round is the same as the path energy fixed point value of the previous round bit by bit; stop when the path binary search round is equal to the number of tree tensor network layers; if no stop is triggered, return to S44 to continue inserting the path midpoint and summing again; after stopping, read the path energy of the stopped round, write zero when the path energy is negative, take the square root when the path energy is zero or positive, generate the twisted Riemann distance, and write it into the twisted Riemann distance table according to the text number and node path code; Through the above processing, the node vectors in the tree attribution results are mapped to potential coordinates. The reconstruction residuals of the clothing ontology tree and the display interference tree change the basic node metrics respectively. The multi-source fusion confidence word further adjusts the path relationship between the text source node and the ontology value node and the interference value node. Finally, a distorted Riemann distance table indexed by text number and node path code is formed, which is used by S5 to perform text semantic vector incorporation, deletion and fusion vector shrinkage. In practical applications: After the waist pattern node of the same dress is written into the clothing ontology tree, the reconstruction residual shortens the path length of the potential position of the waist pattern node. After the background scene node is written into the display interference tree, the reconstruction residual increases the path length of the potential position of the background scene node. After the path energy calculation, the source node of the text "slimming" in the comments is closer to the waist pattern node and farther away from the background scene node. Based on this, S5 writes the semantic vector of the text corresponding to "slimming" into the waist pattern node.

[0021] S5. Based on the distorted Riemann distance, the text semantic vector is written into the first tree node in ascending order of distance. The text semantic vector in the display interference tree is deleted and the interference mapping position is expanded according to the number of deletions. The writing and expansion are repeated according to the tree level. The clothing ontology tree and the matching incremental tree are shrunk to form a style fusion vector. The clothing style perception evaluation result is output through a deep classification neural network. In this embodiment, the text semantic vector is assigned to tree nodes according to the distance fixed point value, and the interference mapping coordinates are rewritten by displaying the deletion feedback of the interference tree, so that the text semantic vector gradually gets rid of the display interference tree in multiple rounds of assignment. Then, the root node style fusion vector is generated by the tree tensor contraction. The distorted Riemann distance table is indexed by the text number and node path code. The tree node belonging word adopts the ontology value, collocation value and interference value in S3. The parent node kernel coefficient is read from the candidate parent node kernel selected in S24. This implementation process includes the following steps: In S51, the semantic vector of a text is written into a single tree node according to the distorted Riemann distance. The attribute word is used to record the attribute position of the text in this round. The input is the distorted Riemann distance table, text number, node path code, region position number, and tree node attribute word. During processing, the distance fixed point value is read in groups according to the text number. Candidate tree nodes are first sorted in ascending order according to the distance fixed point value. If the distance fixed point values ​​are the same, they are rearranged in ascending order according to the node path code. The path code of the first node in the ascending order is taken. The region position number and tree node attribute word corresponding to the path code of the first node are read. The node path code, region position number, and tree node attribute word are concatenated in the order of the fields to form the attribute word, generating the attribute table for this round. When there is no candidate tree node for the same text number, an empty attribute word is written. The text number corresponding to the empty attribute word is not included in the node fusion calculation of S52. In S52, the current round's inclusion table is divided into deletion records and fusion records, so that the semantic vectors of the text in the interference tree are not included in style fusion. The inputs are the current round's inclusion table, the text semantic vector table, the tree node vector table, and the tree node's attribution word. During processing, at the beginning of each round, the deletion count is written to zero according to the region position number. The tree node's attribution word in each inclusion word is read. When the tree node's attribution word is an interference value, the corresponding text number is written to the deletion table, and the deletion count of the same region position number is incremented by one. When the tree node's attribute word is a value taken from the ontology or a combination of words, the corresponding text semantic vector and tree node vector are read, the two are summed bit by bit and then divided by two to generate a node fusion vector, and then written into the node fusion vector table according to the text number, node path code and region number; when the text semantic vector is empty or the tree node vector is empty, the corresponding node fusion vector is written as a zero vector and a fusion missing mark is written. In S53, the deletion count is used to push the interference mapping coordinates away from the body center coordinates, and the next round of incorporation results are formed by recalculating the distorted Riemann distance. The inputs are the deletion count, body mapping coordinates, matching mapping coordinates, interference mapping coordinates, the current round incorporation table and the number of tree layers. During processing, the body mapping coordinates are read for the same region position. If the body mapping coordinates exist, the average value is calculated bit by bit to form the body center coordinates. If the body mapping coordinates do not exist but the matching mapping coordinates exist, the average value of the matching mapping coordinates is read as the body center coordinates. If neither the body mapping coordinates nor the matching mapping coordinates exist, the interference mapping coordinates retain the value of the previous round and are written to the center missing mark. When the body center coordinates exist, the offset vector is obtained by subtracting the body center coordinates from the interference mapping coordinates. The offset vector is multiplied by the result of adding one to the deletion count and then added back to the body center coordinates to obtain the updated interference mapping coordinates. The distorted Riemann distance is recalculated based on the updated interference mapping coordinates and then returned to S51. In the first round, the characters to be assigned in this round are saved and no round comparison is performed. From the second round onwards, the characters to be assigned in this round are compared with the characters to be assigned in the previous round. The process stops when all characters corresponding to the assigned characters are identical. The process stops when the characters to be assigned are not identical and the number of rounds reaches the tree level. The iterative assignment table of the terminated round is output. In S54, the node fusion vector of the terminating round is converted into a style fusion vector through tree tensor shrinkage, and the evaluation result is generated by a deep classification neural network. The inputs are the iterative induction table, the node fusion vector table, the parent node kernel coefficient table, the tree tensor network, and the evaluation dimension table. During processing, the node fusion vectors of the terminating round whose tree node affixation word is an ontology value or a combination value are retained. The child node fusion vectors under the same parent node are read in the direction from the leaf to the root. The child node fusion vectors are multiplied bit by bit by the parent node kernel coefficient and summed. The absolute value sum of the parent node kernel coefficients under the same parent node is calculated. When the absolute value sum of the parent node kernel coefficients is not zero, the summation result is divided by the absolute value sum of the parent node kernel coefficients. When the absolute value sum of the parent node kernel coefficients is zero, a zero vector is written. When there is no node fusion vector under the same parent node, write the zero vector. After shrinking layer by layer to the root, the root node style fusion vector is obtained. Then, the deep classification neural network reads the root node style fusion vector and outputs the classification output value according to the evaluation dimension table. The evaluation dimension number, classification output value, participating shrinkage region number and node path code are written into the clothing style perception evaluation result. After processing by S51 to S54, the text semantic vector is first written into the tree node according to the distorted Riemann distance. The text semantic vector that is included in the interference tree is deleted and the interference mapping coordinate is pushed outward in the opposite direction. The inclusion result stops when the round is stable or the tree layer number is reached. The node fusion vector in the clothing ontology tree and the matching increment tree is fused into a style fusion vector through tree tensor contraction. The style fusion vector is output as a clothing style perception evaluation result by a deep classification neural network. In practical applications: In the first round of distance calculation, the comment phrase "slimming" for the same dress is close to the waist pattern node and the background scene node respectively. If "slimming" is assigned to the background scene node, the count is deleted so that the corresponding interference mapping coordinate of the background scene node moves away from the center coordinate of the waist body. After the next round of distance calculation, "slimming" is written into the waist pattern node. The waist pattern node and the matching belt node jointly participate in the tree tensor shrinkage. Based on this, the deep classification neural network outputs a style-aware evaluation result with the waist region position number and node path encoding.

[0022] Furthermore, the present invention also includes using a dress with the product number "DRS-2026-041" in an e-commerce platform as the clothing object to be evaluated. The platform displays the dress in three display records using an office commuting background, an outdoor leisure background, and a studio meeting background, respectively. The clothing object in the three display records has the same code, and the display order is written as one, two, and three in ascending order of publication time. The first recorded message contained the text "slimming for commuting," was coded for people aged 45 to 55, was tagged as commuting, and had comments including "slimming waistline." The second recorded message contained the text "casual and gentle," was tagged as casual, and had comments including "outerwear doesn't match." The third recorded message contained the text "formal and appropriate," was tagged as meeting, and had comments including "yellowish filter." When processing the first display record, the deep convolutional network scales the clothing image to the training input size and outputs pixel feature maps, human key point coordinates, and clothing contour mask. After the human key point coordinates are horizontally projected onto the boundary of the clothing contour mask, region numbers are generated according to the vertical coordinates of the projection hit points. Among them, region number one corresponds to the collar and shoulder area, region number two corresponds to the waist area, region number three corresponds to the skirt area, and region number four corresponds to the belt area that contacts the waist boundary. The deep convolutional network sums the pixel features covered by each region's position number in each dimension and divides them by the number of covered pixels to obtain the visual vector. Taking the waist region as an example, the visual vector of region position number two in the first display record is [0.72, 0.31, 0.18, 0.09], the visual vector of region position number two in the second display record is [0.70, 0.34, 0.17, 0.11], and the visual vector of region position number two in the third display record is [0.73, 0.32, 0.19, 0.10]. The associated text is processed by a transform coding neural network to output word vectors in word order. The semantic vector of the corresponding word segment for "commuting makes you look thinner" is [0.24, 0.12, 0.36, 0.04]. The age embedding network expands the age layer encoding into an age vector [1, 1, -1, 1]. The word segment semantic vector and the age vector are multiplied bit by bit and then summed to obtain the age inner product of 0.04. When the age inner product is zero or positive, the word segment semantic vector is preserved. The scene embedding network expands the commuting scene into a scene vector, and then multiplies it bit by bit with each visual vector and sums the results. The summation value of region number two is 0.68, which is higher than other region numbers, so "commuting makes you look slimmer" is written into region number two. The bidirectional long short-term memory network reads "waistline makes you look slimmer" and generates a positive sentiment value of 0.81 and a negative sentiment value of 0.13, with a sentiment difference of 0.68. The sentiment value is one, and the display semantic vector [0.24, 0.12, 0.36, 0.04] is written under region number two. For "filter is yellowish", the sentiment value is negative one, and the display semantic vector is written into the corresponding region number in the third display record with a negative value. In S2, the system reads the visual vector and the display semantic vector according to the display order and the region number. If the dimensions are the same, they are added bit by bit to form the region fusion vector. Taking the region number two of the first display record as an example, the visual vector [0.72, 0.31, 0.18, 0.09] and the display semantic vector [0.24, 0.12, 0.36, 0.04] are added bit by bit to obtain the region fusion vector [0.96, 0.43, 0.54, 0.13]. The region fusion vector is written into the feature tensor according to the display order, region number and vector number. The three display orders and four region numbers form a 3x4x4 feature tensor. When the region number is missing, the corresponding tensor cell is written with a zero value and a missing mark. When performing tree tensor network decomposition on feature tensors, region position one and region position two are first merged into the first sub-tensor, and region position three and region position four are merged into the second sub-tensor. The first sub-tensor is expanded along the display order to obtain a three-row eight-column matrix, and the rank enumeration endpoint is three. The system performs singular value decomposition with integer ranks 1, 2, and 3 respectively to generate candidate parent node kernels, and reconstructs the corresponding sub-tensors using the candidate parent node kernels. In the first sub-tensor, the number of residual signs in the same direction corresponding to rank 1 is five, the number of region position numbers is three, and the sort value is eight. The sort value corresponding to rank 2 is seven, and the sort value corresponding to rank 3 is six. Therefore, the candidate parent node kernel corresponding to rank 1 is taken as the first parent node kernel. The second sub-tensor obtains the second parent node kernel in the same way, and the tree tensor network and the parent node kernel coefficient table are formed by shrinking from the leaves to the roots. In S3, the system sets the core of the first parent node to zero according to the parent node order, and the neural decoder reconstructs three display records. The regional residuals under the second region number before and after setting to zero are 0.31, 0.29 and 0.33 respectively. All three regional residuals are positive, the absolute value of the cumulative residual sign bit is three, the number of non-zero regional residuals is three, the region number remains two after reconstruction, the change in region number is zero, the ontology ownership cost is zero, so the first parent node is written into the clothing ontology tree. After the parent node of the corresponding belt area is set to zero, only the first and second display records generate non-zero area residuals. The sign bits of the residuals do not form the same direction of the full display order, but the boundary coordinates of the belt area and the boundary coordinates of the waist area intersect. The occlusion order difference is zero, and the matching belonging cost value is zero. Therefore, the parent node of the belt area is written into the matching incremental tree. After the parent node of the background scenery and filter color shift in the third display record is set to zero, the area position number is migrated, and there is no boundary continuation with the clothing body tree node. Therefore, it is written into the display interference tree. In S4, the neural encoder maps the waist node in the clothing ontology tree to latent coordinates [0.18, 0.24] and the background scenery node in the display interference tree to latent coordinates [0.86, 0.79]. The neural decoder generates a Jacobian matrix at each latent coordinate and multiplies the Jacobian matrix by the transpose matrix and adds one to the main diagonal elements to form the node's basic metric. In region 2, the sum of the ontology residuals is 0.93, the sum of the interference residuals is 0.20, the shrinkage factor is (0.93+1) / (0.93+0.20+2), and the expansion factor is (0.93+0.20+2) / (0.20+1). The basic metric of the ontology value node is shrunk, and the basic metric of the interference value node is expanded. The path energy from the text source node "Slim" to the waist body node is 0.14, and the interference path energy to the background interference node is 1.37. The path energy difference is negative, the body support count is incremented by one, and the multi-source fusion confidence word is written into the node metric table. After path bisection and energy summation, the distorted Riemann distance from the text number corresponding to "Slim" to the waist body node is 0.37, and the distorted Riemann distance to the background interference node is 1.16. In S5, the system generates the classifying words based on the distorted Riemann distance table according to the text number; the first digit of the ascending distance of the text number corresponding to "slimming" is the waist body node, and the classifying word is formed by concatenating the path code of the waist body node, the second region position number, and the body value; the text number corresponding to "outerwear doesn't match" is classified into the belt matching incremental node; the text number corresponding to "filter is yellowish" is classified into the display interference tree node, and the system writes "filter is yellowish" into the deletion table and increments the corresponding region deletion count by one; Subsequently, the system subtracts the body center coordinates from the interference mapping coordinates to form an offset vector. The offset vector is multiplied by the deletion count plus one and then added back to the body center coordinates. The interference mapping coordinates are updated and the distorted Riemann distance is recalculated. The second round of input words are identical to the first round of input words, and the input process stops. In the final round, the clothing ontology tree retains the node fusion vectors of the waist ontology node and the collar and shoulder ontology node, and the incremental tree retains the node fusion vectors of the belt node, showing that the semantic vectors of the text related to "filter yellowing" and background setting in the interference tree do not enter the shrinking process. The system reads the parent node kernel coefficients from the leaves to the root, multiplies the node fusion vectors under the same parent node by the parent node kernel coefficients bit by bit and sums them, and then divides them by the sum of the absolute values ​​of the parent node kernel coefficients to obtain the root node style fusion vector [0.61, 0.48, 0.53, 0.27]. After reading the style fusion vector from the root node, the deep classification neural network outputs the clothing style perception evaluation results according to the evaluation dimension table: fashionability output value is 0.82, age suitability output value is 0.76, scene matching output value is 0.74, aesthetic acceptance output value is 0.80, and the region number involved in the output includes region number one, region number two, and region number four. The background scene node path code corresponding to the interference tree is not written into the participating region field. In this implementation method, the display records of the same dress under different models, backgrounds, filters, and outfits are included in the same processing process. The waist shape and neckline and shoulder contours are entered into the garment body tree because they are stable across the display record residual direction. The belt is entered into the outfit increment tree because it is connected to the waist boundary and maintains the occlusion order. The background setting and filter color deviation are entered into the display interference tree because of the migration of the area number and the lack of boundary continuity. The final output garment style perception evaluation result is supported by the garment body area and the actual outfit relationship.

[0023] Working principle: This solution first uses edge computing nodes to process multiple display records of the same garment together, instead of just looking at a single image and text content. The system first extracts visual vectors of different garment regions from the garment image, and then converts the matching text, age group, scene tags, and comment sentiment into semantic vectors that can be written into the corresponding garment regions. Then, these regional features are combined into feature tensors according to display order and region position, and decomposed into parent node structures through a tree tensor network. The system then sets each parent node to zero and reconstructs the display record on the edge computing side. Based on the residual direction before and after reconstruction, whether the region migrates, whether the boundary is continuous, and whether the occlusion order is maintained, the features are divided into garment body, matching increment, and display interference. Then, through distorted Riemann metric learning, the distance between the text and the garment body or matching relationship is shortened, while the distance between the text and display interference is widened. Finally, only the garment body and the actual matching relationship are used to form a style fusion vector, and the garment style perception evaluation result is output. For example, the same dress might be displayed on an e-commerce platform in different settings: an office, outdoors, and a studio. The background, filters, model poses, and belts all differ, and comments might mention things like "slimming," "formal," or "yellowish filter." Edge computing nodes would first extract areas like the neckline, waist, hem, and belt, and write descriptions like "slimming for commuting" into the corresponding areas. If the waistline maintains the same shape across multiple displays, it's categorized as part of the garment itself. If the belt connects to the waistline and maintains the same occlusion order, it's categorized as part of the styling increment. If the filter is yellowish or the background only changes with the shooting conditions, it's categorized as display interference and excluded from the reviews. In this way, the output's fashionability, age suitability, scene matching, and aesthetic appeal are no longer skewed by the background and filters, but are primarily supported by the garment's design and the actual wearing relationship identified by edge computing.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating clothing style perception based on multimodal feature fusion, characterized in that, include: S1. Obtain the display records of the same clothing object, use deep learning to extract visual vectors with region identifiers from the clothing images, encode the text into text semantic vectors, adjust the direction by age group and write it into the region identifier pointed to by the scene label, and rewrite the value according to the comment sentiment symbol. S2. Stack visual vectors according to the display order and embed text semantic vectors to form feature tensors and perform tree tensor network decomposition. With the display order as the root and the region position as the leaf, enumerate the sub-tensor ranks and perform singular value decomposition. Determine the shrinking rank according to the sum of the number of identical signs in the reconstruction residual and the number of region position, and generate a tree tensor network. S3. Set each parent node to zero and reconstruct the display record through the neural decoder. Write the clothing body tree when the residual sign and absolute value are the same as the non-zero residual number. Otherwise, write the matching increment tree when the parent node area is connected to the boundary of the clothing body tree node area and the occlusion order remains unchanged. Write the remaining parent nodes into the display interference tree. S4. Perform Twisted Riemann metric learning on the tree nodes. After mapping by the neural encoder, multiply the Jacobi matrix of the neural decoder by the transpose matrix to generate the pulled-back Riemann metric. Based on the reconstruction residuals of the clothing ontology tree and the display interference tree, shrink the ontology mapping position and expand the interference mapping position. Generate the twisted Riemann distance through path bisection and energy summation. S5. Based on the distorted Riemann distance, the text semantic vector is written into the first tree node in ascending order of distance. The text semantic vectors in the display interference tree are deleted and the interference mapping position is expanded according to the number of deletions. The writing and expansion are repeated according to the tree level. The clothing ontology tree and the matching increment tree are shrunk to form a style fusion vector. The clothing style perception evaluation result is output through a deep classification neural network.

2. The clothing style perception evaluation method based on multimodal feature fusion according to claim 1, characterized in that: S1 includes: S11. For a single display record with the same clothing object encoding, the deep convolutional network reads the clothing image and generates a pixel feature map. The human body key point branch outputs the key point coordinates according to the human skeleton order, and the clothing segmentation branch outputs the clothing contour mask. The key point coordinates are projected horizontally onto the boundary of the clothing contour mask. The region number is generated in ascending order according to the vertical coordinate of the projection hit point. Then, the pixel features covered by the region number are summed dimension by dimension and divided by the number of covered pixels to output a visual vector with region number. S12. For related text in a single display record, the transform coding neural network outputs word vectors in word order, generates a segment semantic vector by the average of word vectors between sentence-end markers, and the age embedding network expands the age layer encoding into an age vector. The segment semantic vector and the age vector are multiplied bit by bit and then summed to obtain the age inner product. When the age inner product is negative, the segment semantic vector is inverted bit by bit. When the age inner product is zero or positive, the segment semantic vector is preserved, and the age-adjusted semantic vector is output. S13. The scene embedding network expands the scene label into a scene vector. The scene vector is multiplied bit by bit by each visual vector and then summed. The sum is sorted in descending order and in ascending order by region number when the sums are the same. The bidirectional long short-term memory network reads the comment segment and generates a sentiment difference through a linear layer. When the sentiment difference is negative, it is taken as negative one, and when the sentiment difference is zero or positive, it is taken as one. The value is multiplied into the age-adjusted semantic vector and then written into the region number. This generates a display semantic vector that is aligned and fused with the visual vector according to the region number.

3. The clothing style perception evaluation method based on multimodal feature fusion according to claim 2, characterized in that: S2 includes: S21. For each region number under the display order, read the visual vector and text semantic vector corresponding to the region number. When the dimensions are inconsistent, pad the tail of the short-dimensional vector with zeros until the dimensions are consistent. Then add them one by one and record the region number to generate the region fusion vector. S22. Using the display order as the first dimension, the region position number as the second dimension, and the vector position number as the third dimension, write the region fusion vector into the corresponding tensor cell. Write zero values ​​into the tensor cells that have not obtained the region fusion vector and mark them as missing, thus generating the feature tensor.

4. The clothing style perception evaluation method based on multimodal feature fusion according to claim 3, characterized in that: S2 also includes: S23. Perform tree tensor network decomposition on the feature tensor to connect the root of the tree in the display order and connect the leaves with the region position number. Merge the sub-tensors according to the adjacent region position numbers. Expand each sub-tensor into a matrix along the display order. When the number of rows in the matrix does not exceed the number of columns, use the number of rows as the rank enumeration endpoint. When the number of rows in the matrix exceeds the number of columns, use the number of columns as the rank enumeration endpoint. Perform singular value decomposition on each integer rank between the first and the rank enumeration endpoint to generate candidate parent node kernels. S24. Reconstruct the corresponding sub-tensor using candidate parent node kernels. Subtract the original tensor value from the reconstructed value to form the reconstruction residual. Take 1 when the reconstruction residual is positive, take -1 when it is negative, and take 0 when it is zero. Sort the candidate parent node kernels in descending order according to the sum of the number of residuals with the same sign in the same region and the number of regions with the same sign after reconstruction. When they are in the same position, take the first candidate parent node kernel in ascending order of integer rank to shrink the sub-tensor. Shrink the sub-tensor layer by layer from the leaves to the root and output the tree tensor network.

5. The clothing style perception evaluation method based on multimodal feature fusion according to claim 4, characterized in that: S3 includes: S31. For the parent node in the tree tensor network, set the parent node kernel to zero according to the parent node order and reconstruct and display the record through the neural decoder. Read the difference of the fusion vector under the same region position before and after setting it to zero. Summate the fusion vector difference dimensionally to obtain the region residual. Write 1 if the region residual is positive, write negative 1 if it is negative, and write zero if it is zero. Generate the residual sign bit and record the region position and occlusion order after reconstruction. S32. Based on the residual sign bit of the same parent node in all displayed records, count the number of non-zero region residuals, the absolute number of the cumulative residual sign bit value, and the change in region position number. Generate the ontology attribution value by subtracting the absolute number of the cumulative residual sign bit value from the number of non-zero region residuals and adding the change in region position number.

6. The clothing style perception evaluation method based on multimodal feature fusion according to claim 5, characterized in that: S3 also includes: S33. Generate the initial node of the clothing ontology tree from the parent node with the ontology attribution cost value of zero. For the parent node with the ontology attribution cost value of non-zero, calculate the number of the same coordinates between its region boundary coordinates and the region boundary coordinates of the initial node of the clothing ontology tree. Determine the successor node in descending order of the number of the same coordinates. Form the boundary missing position with the number of the same coordinates of zero and form the occlusion offset value with the absolute value of the occlusion order difference. Add the boundary missing position and the occlusion offset value to generate the matching attribution cost value. S34. Take the sum of the attribution costs of all parent nodes as the first target bit and the number of matching attributions without successor nodes as the second target bit. Determine the attribution of parent nodes by the first position in ascending order of the target words. Among them, the parent nodes with the attribution cost of the main body are written into the clothing main body tree, the parent nodes with the attribution cost of the matching are written into the matching incremental tree, and the remaining parent nodes are written into the display interference tree. The first round takes the attribution result of the previous round as all zeros as input. When the attribution result is XORed with the attribution result of the previous round node by node and the sum is zero, the tree attribution result is output. Otherwise, the number of the same coordinates and the occlusion order difference are recalculated to the tree level based on the attribution result.

7. The clothing style perception evaluation method based on multimodal feature fusion according to claim 6, characterized in that: S4 includes: S41. Based on the tree attribution results, read the node vector, region number, source position and reconstruction residual of each tree node. The neural encoder maps the node vector to the latent coordinates. The neural decoder generates the Jacobian matrix at the latent coordinates and multiplies the Jacobian matrix by the transpose matrix and adds one to the diagonal to form the basic metric of the node. S42. For the same region reference number, sum the absolute values ​​of the reconstruction residuals where the node's attribute word is the ontology value to obtain the ontology residual sum, and sum the absolute values ​​of the reconstruction residuals where the node's attribute word is the interference value to obtain the interference residual sum. Add one to the ontology residual sum and divide by the ontology residual sum plus the interference residual sum plus two to form a contraction factor. Add two to the ontology residual sum plus the interference residual sum and divide by the interference residual sum plus one to form an expansion factor. Use the contraction factor to rewrite the basic metric of the ontology value node, and use the expansion factor to rewrite the basic metric of the interference value node.

8. The clothing style perception evaluation method based on multimodal feature fusion according to claim 7, characterized in that: S4 also includes: S43. For visual source nodes and text source nodes in the same region, calculate the ontology path energy from the text source node to the ontology value node and the interference path energy from the text source node to the interference value node under the rewritten node base metric. Subtract the interference path energy from the ontology path energy to form the path energy difference. When the path energy difference is negative, increment the ontology support count by one. When the path energy difference is zero, increment both the ontology support count and the interference support count by one. When the path energy difference is positive, increment the interference support count by one. Then, divide the incremented ontology support count by the sum of the ontology support count and the interference support count by two to generate the multi-source fusion confidence word. S44. Multiply the multi-source fusion confidence word into the basic metric of the ontology value node, multiply the result of subtracting the multi-source fusion confidence word from one and then adding one into the basic metric of the interference value node, perform path bisection on the adjacent potential coordinates according to the tree edge connection order, take the average of the two potential coordinates at the midpoint, and obtain the path energy of this round by the product of the adjacent coordinate difference vector, the node basic metric and the transpose of the adjacent coordinate difference vector. S45. Record the path energy in the first round. Starting from the second round, stop the distorted Riemann metric learning when the fixed value of the path energy in the current round is the same as the fixed value of the path energy in the previous round, or when the number of path bisection rounds is equal to the number of tree tensor network layers. The square root of the path energy in the current round is used as the distorted Riemann distance output.

9. The clothing style perception evaluation method based on multimodal feature fusion according to claim 8, characterized in that: S5 includes: S51. Based on the distorted Riemann distance table, read the distance fixed point value in groups according to the text number, rearrange the candidate tree nodes with the same distance fixed point value in ascending order according to the node path code, take the path code of the first node in ascending order, the corresponding region number and the tree node belonging word to concatenate the classification word, and generate the current round of classification table. S52. Read the tree node's belonging word according to the current round's inclusion table. When the belonging word is an interference value, write the corresponding text number into the deletion table and increment the deletion count in the same area. When the belonging word is a non-interference value, sum the text semantic vector and the tree node vector bit by bit and divide by two to generate the node fusion vector.

10. The clothing style perception evaluation method based on multimodal feature fusion according to claim 9, characterized in that: S5 also includes: S53. Update the interference mapping coordinates based on the deletion count. First, calculate the average value of the body mapping coordinates in the same region to form the body center coordinates. Then, subtract the body center coordinates from the interference mapping coordinates to obtain the offset vector. Multiply the offset vector by the deletion count plus one and add it back to the body center coordinates. Recalculate the distorted Riemann distance and return to S51. Stop when each of the input words is equal to the input word of the previous round. If they are not the same, stop when the number of rounds reaches the tree level. Output the iterative input table. S54. Based on the iterative induction table, retain the node fusion vector of the final round, perform tree tensor contraction from the leaves to the root, multiply the child node fusion vector under the same parent node bit by the parent node kernel coefficient and sum them. When the absolute value of the coefficient is non-zero, divide it by the sum of the absolute values ​​of the coefficients. When the sum of the absolute values ​​of the coefficients is zero, write the zero vector to obtain the root node style fusion vector. Then, perform classification-level fusion by the deep classification neural network and output the clothing style perception evaluation result.

Images