Clothing compatibility analysis method and equipment based on multi-modal feature fusion and medium

Abstract

The invention relates to a clothing compatibility analysis method and equipment based on multi-modal feature fusion and a medium, and the method comprises the steps: carrying out the feature extraction of an input suit picture, and obtaining the semantic feature, visual feature and attribute feature of each single item in a suit; and processing the attribute features by using the attribute masks to obtain compatibility features of the attribute levels. Inputting the semantic feature, the attribute feature and the visual feature of each single item in the suit into a full connection layer, and fusing to obtain a compatibility feature of a visual level and a compatibility feature of a cross-modal level; obtaining a garment compatibility score of each attribute of each single item of the suit and a garment compatibility score of each single item at each level by using a multi-level graph convolutional network; and calculating a total clothing compatibility score to obtain the importance degree of each attribute in the suit to the suit compatibility and the importance degree of each single item to the suit compatibility, and outputting a compatibility description result. Compared with the prior art, the method has the advantages of multiple modes, high interpretability, high precision and the like.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and image processing, and in particular to a clothing compatibility analysis method, device and medium based on multimodal feature fusion. Background Art

[0002] Clothing compatibility analysis involves analyzing a set of clothing images based on the implicit compatibility information within them to determine whether the garments are compatible. Traditionally, this task required manual analysis by professionals such as fashion designers, which was labor-intensive and time-consuming. In recent years, the development of deep learning technology has streamlined the traditional analysis process, leading to the emergence of automated methods based on neural networks. This allows ordinary users without specialized knowledge to quickly obtain compatibility analysis results by simply inputting clothing imagery into the model.

[0003] With economic development and the booming fashion market, consumers are increasingly interested in fashion styles and matching effects. Accordingly, clothing compatibility analysis has become a research hotspot in the field of computer vision. Numerous researchers have devoted themselves to this area, employing diverse methodological approaches. From a relational structure perspective, existing research can be categorized into three main categories: compatibility analysis methods based on clothing item pairing, compatibility analysis methods based on clothing item sequences, and compatibility analysis methods based on graph structures. Although existing methods have made some progress in clothing compatibility analysis, they still have significant shortcomings. First, the single feature modality leads to insufficient accuracy. Existing methods typically extract and analyze the compatibility relationships of fashion outfits from a single modality (e.g., visual features or attribute features alone), ignoring the diversity and richness of compatibility information itself. This single-modality limitation makes it difficult for models to comprehensively and accurately capture complex compatibility relationships, resulting in suboptimal judgment accuracy. Second, the poor interpretability of the results limits their guiding significance. Existing methods often output simple numerical scores (e.g., a scalar between 0 and 1), which can only represent a binary conclusion of "compatible" or "incompatible." This score-based result lacks specific explanations, and cannot tell users why they are compatible or incompatible, which parts have problems, or which are properly matched. Therefore, the results obtained by existing methods are not very helpful for users' actual matching and cannot meet the application needs in real scenarios.

[0004] Chinese patent application CN117635769A discloses a method for recommending clothing in social networks based on an association-aware cross-modal attention network. This method enriches the feature representation of fashion items by fusing cross-modal information and propagating the contextual relationships of individual items, ensuring the consistency and complementarity of multimodal information, thereby achieving the task of recommending complementary clothing items. However, the clothing recommendations in this application are based on a compatibility score, which can only determine whether the items are compatible or not, and cannot provide specific reasons and analysis. Therefore, how to overcome the problems of insufficient judgment accuracy caused by the single feature modality in the existing technology, and the limited user guidance significance caused by the poor interpretability of the results, so as to obtain a more accurate and more practical clothing compatibility analysis method, is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a clothing compatibility analysis method, equipment and medium based on multimodal feature fusion. Through a multi-level graph convolutional network, the compatibility feature analysis of each attribute of each item is performed to obtain the compatibility features of the attribute level, the compatibility features of the visual level and the compatibility features of the cross-modal level, and then obtain the total compatibility score of the suit, the importance of each attribute to the compatibility of the suit and the importance of each item in the suit to the compatibility of the suit, thereby realizing the generation of compatibility description and obtaining richer and more complete clothing compatibility information.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to one aspect of the present invention, a clothing compatibility analysis method based on multimodal feature fusion is provided, which specifically comprises the following steps:

[0008] S1. Multi-level feature extraction: extract features from the input set image using a pre-trained feature extraction network to obtain the semantic and visual features of each item in the set. The visual features are then decoupled using a multi-layer perceptron to obtain attribute features.

[0009] S2. Perform multimodal feature fusion through a multi-level graph convolutional network, process attribute features using attribute masks, and obtain attribute-level compatibility features; input the attribute features and visual features of each item in the set into a fully connected layer, and fuse them to obtain visual-level compatibility features; input the visual features and semantic features of each item in the set into a fully connected layer, and fuse them to obtain cross-modal-level compatibility features; the multi-level graph convolutional network includes a visual-level graph convolutional network, a cross-modal-level graph convolutional network, and an attribute-level graph convolutional network;

[0010] S3. Using a multi-level graph convolutional network, based on the attribute-level compatibility features, the visual-level compatibility features, and the cross-modal-level compatibility features, obtain the clothing compatibility scores of each item in the suit at each level.

[0011] S4. Generate compatibility analysis results. The multi-layer perceptron obtains the total clothing compatibility score based on the clothing compatibility scores of the fashion suit at each level, calculates the importance of each attribute in the suit to the suit compatibility, and calculates the importance of each single item to the suit compatibility. The compatibility description result is obtained through the visual language model.

[0012] Furthermore, in S1, the set image is decomposed into several items to obtain visual features, and the visual features of each item are decoupled using a multi-layer perceptron with the same number of attributes according to several predefined attributes to obtain attribute features.

[0013] Furthermore, each node of the multi-level graph convolutional network corresponds to a single item in the set, and each edge between the nodes represents the compatibility relationship between the single items corresponding to the nodes.

[0014] Furthermore, the initialization of the attribute level graph convolutional network in S2 specifically includes: during the learning process of the attribute level graph convolutional network, each node With a hidden state vector When a single item has a corresponding attribute, the hidden state vector is initialized to the corresponding attribute feature, otherwise it is initialized to an all-zero vector. The expression is:

[0015]

[0016] in, is the mask in training, when the k attributes of the i-th item are meaningful otherwise is the attribute feature corresponding to the k attribute of the i-th item, and K is the number of multi-layer perceptrons.

[0017] Furthermore, the hidden state vector g of the i-th node in the hierarchical graph convolutional network in S2 is i Initialization, based on the visual features v of a single product i and the updated hidden state vector of the node in the attribute-level graph convolutional network The expression is:

[0018]

[0019] Among them, [·||·] represents vector splicing, W h is the weight matrix to be learned for the fully connected layer that propagates from the attribute level to the visual level.

[0020] Furthermore, the hidden state vector c of the i-th node in the cross-modal hierarchical graph convolutional network in S2 is i Initialization, based on the semantic features of the single product e i and the updated hidden state vector of the node in the visual hierarchy graph convolutional network The expression is:

[0021]

[0022] Where W g It is the weight matrix to be learned by the fully connected layer from the visual layer to the cross-modal layer.

[0023] Furthermore, in S4, the compatibility scores of each item in each attribute in the attribute hierarchy are weighted and summed to obtain the importance of each attribute to the compatibility of the set, which is expressed as follows:

[0024]

[0025] Among them, S k is the importance of each attribute k to the compatibility of the package, W i is the preset weight of the i-th item, is the compatibility of attribute k of the i-th item, and s is the total clothing compatibility score.

[0026] Furthermore, in S4, a weighted sum of the compatibility scores of the attribute level, visual level, and cross-modal level is performed on each single item to obtain the importance of each single item in the set to the set compatibility, which is expressed as:

[0027]

[0028] Among them, S i is the importance of the i-th item in the set to the compatibility of the set, W l is the predefined level weight, L is the number of attributes, s is the total clothing compatibility score, is the compatibility score of attribute l of the i-th item.

[0029] According to a second aspect of the present invention, an electronic device is provided, comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the method when executing the program.

[0030] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method described above is implemented.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) Multi-level feature fusion improves the accuracy of compatibility analysis: Through the three-level progressive feature fusion architecture of attribute → vision → cross-modality, the decoupled attribute features are first masked to generate attribute-level compatibility features; then the attribute features and visual features are fused through the fully connected layer to generate visual-level compatibility features; finally, the visual features and semantic features are fused to generate cross-modal compatibility features, which overcomes the limitations of single-modality analysis and realizes multi-level compatibility relationship modeling from fine-grained attributes to high-level semantics, significantly improving the judgment accuracy.

[0033] (2) Fine-grained attribute decoupling enhances feature expression capabilities: K independent MLPs are used to decouple visual features, explicitly extracting multi-dimensional attribute features within a single product, such as material and style, and combining them with K parallel attribute hierarchical graph networks to construct a unique compatibility relationship network for each attribute. This design breaks through the limitations of traditional coarse-grained category labels and achieves precise positioning of attribute-level compatibility conflicts.

[0034] (3) Explainable compatibility descriptions guide actual pairing: Attribute importance and item importance are calculated through a regularized weighting mechanism, and natural language descriptions are generated in conjunction with a visual language model. This design transforms abstract compatibility scores into specific analyses, such as "Material conflict: The breathability of a cotton top and polyester bottoms does not match," addressing the lack of guidance provided by existing technical results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of the clothing compatibility analysis method based on multimodal feature fusion;

[0036] Figure 2 This is the data flow diagram of the clothing compatibility analysis method based on multimodal feature fusion. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0038] like Figure 1 and Figure 2 As shown, this embodiment provides a clothing compatibility analysis method based on multimodal feature fusion, and the specific steps include:

[0039] S1. Multi-level feature extraction: extract features from the input set image using a pre-trained feature extraction network to obtain the semantic and visual features of each item in the set. The visual features are then decoupled using a multi-layer perceptron to obtain attribute features.

[0040] S2. Perform multimodal feature fusion through a multi-level graph convolutional network, process attribute features using attribute masks, and obtain attribute-level compatibility features. Input the attribute features and visual features of each item in the set into a fully connected layer, and fuse them to obtain visual-level compatibility features. Input the visual features and semantic features of each item in the set into a fully connected layer, and fuse them to obtain cross-modal-level compatibility features. The multi-level graph convolutional network includes a visual-level graph convolutional network, a cross-modal-level graph convolutional network, and an attribute-level graph convolutional network.

[0041] S3. Based on the compatibility features at the attribute level, the visual level, and the cross-modal level, a multi-level graph convolutional network is used to obtain the clothing compatibility scores of each item in the suit at each level.

[0042] S4. Generate compatibility analysis results. The multi-layer perceptron obtains the total clothing compatibility score based on the clothing compatibility scores of the fashion suit at each level, calculates the importance of each attribute in the suit to the suit compatibility, and calculates the importance of each single item to the suit compatibility. The compatibility description result is obtained through the visual language model.

[0043] In S1, the set image is decomposed into several items to obtain the visual features v i =CNN(V i ), where V i is the actual visual image of the i-th item, is the visual feature extracted from the i-th item, D v is the dimension of the visual feature. Semantic feature extraction is achieved through the pre-trained cross-modal basic model OpenFashionCLIP. Visual feature extraction is achieved by extracting features from clothing item images using the pre-trained convolutional neural network ResNet-18. Based on a number of pre-defined attributes, the visual features of each item are decoupled using a multi-layer perceptron with the same number of attributes to obtain attribute features. in is the kth attribute feature of the i-th item, the extracted visual feature, D f is the dimension of the attribute feature.

[0044] Each node in the multi-level graph convolutional network corresponds to an item in a set, and each edge between nodes represents the compatibility relationship between the items corresponding to the nodes. The initial representation of nodes in the attribute-level graph network comes from the attribute-level compatibility features from the previous step; similarly, the initial representation of nodes in the visual-level graph network comes from the visual-level compatibility features from the previous step; and the initial representation of nodes in the cross-modal hierarchical graph network comes from the cross-modal hierarchical compatibility features from the previous step.

[0045] The initialization of the attribute level graph convolutional network in S2 specifically includes: During the learning process of the attribute level graph convolutional network, each node With a hidden state vector Association, when a single item has a corresponding attribute, the hidden state vector is initialized to the corresponding attribute feature, otherwise it is initialized to an all-zero vector. The expression is:

[0046]

[0047] in, is the mask in training, when the k attributes of the i-th item are meaningful otherwise is the attribute feature corresponding to the k attribute of the i-th item, and K is the number of multi-layer perceptrons.

[0048] The hidden state vector g of the i-th node in the hierarchical graph convolutional network in S2 i Initialization, based on the visual features v of a single product i and the updated hidden state vector of the node in the attribute-level graph convolutional network The expression is:

[0049]

[0050] Among them, [·||·] represents vector splicing, W h is the weight matrix to be learned for the fully connected layer that propagates from the attribute level to the visual level.

[0051] The hidden state vector c of the i-th node in the cross-modal hierarchical graph convolutional network in S2 i Initialization, based on the semantic features of the single product e i and the updated hidden state vector of the node in the visual hierarchy graph convolutional network The expression is:

[0052]

[0053] Where W g It is the weight matrix to be learned by the fully connected layer from the visual layer to the cross-modal layer.

[0054] In S4, the compatibility scores of each item in each attribute in the attribute hierarchy are weighted and summed to obtain the importance of each attribute to the compatibility of the set, which is expressed as:

[0055]

[0056] Among them, S k is the importance of each attribute k to the compatibility of the package, W iis the preset weight of the i-th item, is the compatibility of attribute k of the i-th item, s is the total clothing compatibility score. i Usually, the weight of main items such as tops and bottoms is set to 1 to 0.7, and the weight of secondary items such as watches and earrings is set to 0.5 to 0.1.

[0057] In S4, the weighted sum of the compatibility scores of the attribute level, visual level, and cross-modal level is performed on each single item to obtain the importance of each single item in the set to the set compatibility, which is expressed as:

[0058]

[0059] Among them, S i is the importance of the i-th item in the set to the compatibility of the set, W l is the predefined level weight, which defaults to 1, L is the number of attributes, s is the total clothing compatibility score, is the compatibility score of attribute l of the i-th item.

[0060] The total compatibility score is: Where W3, W4, b3 and b4 are MLP parameters, It is obtained by connecting the compatibility scores s1, s2, and s3 at the attribute level, visual level, and cross-modal level.

[0061] After obtaining the importance of each attribute and each item through these processes, the compatibility data is input into the visual language model together with the suit image, and the information is fully utilized through the prompting project. The specific prompt template used in this embodiment is: This is a set of clothing with a compatibility score of s. As an expert in the field of clothing design, please analyze why this set of clothing is compatible / incompatible. Note that the specific importance of each attribute in the suit is S. k , the importance of each item in the set is S i , it is necessary to explain the specific attributes or reasons why a specific item is compatible / incompatible. Finally, the image understanding and text generation capabilities of the visual language model are used to generate the final compatibility description result of the method of the present invention.

[0062] This embodiment introduces multimodal feature fusion in clothing compatibility analysis. For an input clothing set image, it utilizes multiple pre-trained feature extraction networks and feature decoupling to perform multi-level feature extraction. In the next step, multimodal feature fusion is performed, summing the features at each level according to the weight matrix to be learned. Finally, the fused features are fed into the graph convolutional network at the corresponding level for training. Compared with existing single-modality methods, multimodal feature fusion obtains richer and more complete clothing compatibility information. Furthermore, by constructing a multi-level graph convolutional network, it fully utilizes compatibility information at each level for clothing compatibility analysis. This effectively addresses the issues of ambiguous results and inaccurate judgments in compatibility analysis tasks, improving the accuracy of clothing compatibility analysis.

[0063] Furthermore, this embodiment implements compatibility description generation during clothing compatibility analysis. For the compatibility score obtained through clothing compatibility analysis, key data is filtered and processed according to predefined rules and fed into a visual language model along with images of clothing ensembles. A detailed compatibility description is generated through prompting engineering. Compared to existing clothing compatibility analysis methods, this invention generates detailed compatibility descriptions rather than simple compatibility scores ranging from 0 to 1, improving the interpretability of compatibility results. This effectively addresses the problem that the scores generated by existing compatibility analysis methods are insufficiently informative for users and difficult to apply in practice.

[0064] Table 1 compares the experimental results of this embodiment's method with existing compatibility analysis methods. As can be seen from the table, compared with other compatibility analysis methods, this embodiment achieves higher accuracy in determining the compatibility of fashion suits, more accurately and comprehensively captures clothing compatibility information, and generates more interpretable compatibility descriptions.

[0065] Table 1 Comparison of experimental results between the method of this embodiment and the existing compatibility analysis method

[0066] OCP(AUC) FITB(ACC) Result form PS-OCM 0.9009 0.5412 Simple fractions DMAP 0.9156 0.5813 Simple fractions Method of the present invention 0.9266 0.6186 Specific compatibility description

[0067] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0068] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0069] Multiple components in the device are connected to the I / O interface, including: input units, such as a keyboard, mouse, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as magnetic disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device via ROM and / or the communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method of the present invention described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the method of the present invention by any other suitable means (e.g., by means of firmware).

[0070] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0071] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0072] In the context of the present invention, machine-readable medium can be a tangible medium that can contain or store a program for use with an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. Machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. More specific examples of machine-readable storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A clothing compatibility analysis method based on multimodal feature fusion, characterized in that: The specific steps include: S1. Multi-level feature extraction: extract features from the input set image using a pre-trained feature extraction network to obtain the semantic and visual features of each item in the set. The visual features are then decoupled using a multi-layer perceptron to obtain attribute features. S2. Perform multimodal feature fusion through a multi-level graph convolutional network, process attribute features using attribute masks, and obtain attribute-level compatibility features; input the attribute features and visual features of each item in the set into a fully connected layer, and fuse them to obtain visual-level compatibility features; input the visual features and semantic features of each item in the set into a fully connected layer, and fuse them to obtain cross-modal-level compatibility features; the multi-level graph convolutional network includes a visual-level graph convolutional network, a cross-modal-level graph convolutional network, and an attribute-level graph convolutional network; S3. Based on the attribute-level compatibility features, the visual-level compatibility features, and the cross-modal-level compatibility features, a multi-level graph convolutional network is used to obtain the clothing compatibility scores of each attribute of each item in the suit, and the clothing compatibility scores of each item at each level. S4. Generate compatibility analysis results. The multi-layer perceptron obtains the total clothing compatibility score based on the clothing compatibility scores of the fashion suit at each level, calculates the importance of each attribute in the suit to the suit compatibility, and calculates the importance of each single item to the suit compatibility. The compatibility description result is obtained through the visual language model.

2. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: In S1, the set image is decomposed into several items to obtain visual features. According to several predefined attributes, a multi-layer perceptron with the same number of attributes is used to decouple the visual features of each item to obtain attribute features.

3. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: Each node of the multi-level graph convolutional network corresponds to a single item in the set, and each edge between the nodes represents the compatibility relationship between the single items corresponding to the nodes.

4. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: The initialization of the attribute level graph convolutional network in S2 specifically includes: during the learning process of the attribute level graph convolutional network, each node With a hidden state vector When a single item has a corresponding attribute, the hidden state vector is initialized to the corresponding attribute feature, otherwise it is initialized to an all-zero vector. The expression is: in, is the mask in training, when the k attributes of the i-th item are meaningful otherwise is the attribute feature corresponding to the k attribute of the i-th item, and K is the number of multi-layer perceptrons.

5. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: The hidden state vector g of the i-th node in the hierarchical graph convolutional network in S2 i Initialization, based on the visual features v of a single product i and the updated hidden state vector of the node in the attribute-level graph convolutional network The expression is: Among them, [·||·] represents vector splicing, W h is the weight matrix to be learned for the fully connected layer that propagates from the attribute level to the visual level.

6. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: The hidden state vector c of the i-th node in the cross-modal hierarchical graph convolutional network in S2 i Initialization, based on the semantic features of the single product e i and the updated hidden state vector of the node in the visual hierarchy graph convolutional network The expression is: Where W g It is the weight matrix to be learned by the fully connected layer from the visual layer to the cross-modal layer.

7. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: In S4, the compatibility scores of each item in each attribute in the attribute hierarchy are weighted and summed to obtain the importance of each attribute to the compatibility of the set, which is expressed as follows: Among them, S k is the importance of each attribute k to the compatibility of the package, W i is the preset weight of the i-th item, is the compatibility of attribute k of the i-th item, and s is the total clothing compatibility score.

8. The clothing compatibility analysis method based on multimodal feature fusion according to claim 1 is characterized in that: In S4, the weighted sum of the compatibility scores of the attribute level, visual level, and cross-modal level is performed on each single item to obtain the importance of each single item in the set to the set compatibility, which is expressed as: Among them, S i is the importance of the i-th item in the set to the compatibility of the set, W l is the predefined level weight, L is the number of attributes, s is the total clothing compatibility score, is the compatibility score of attribute l of the i-th item.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images