A cross-modal small group pedestrian re-identification method based on optimal subset learning

Abstract

This invention discloses a cross-modal small-group pedestrian re-identification method based on optimal subset learning, comprising: constructing a feature extraction network; introducing a two-dimensional global refinement mechanism during the feature extraction process of the backbone network to obtain an enhanced set of individual features that are insensitive to modal differences; in the training and optimization phase of the network, using cross-modal intra-group consistency constraints to guide the feature distribution, and combining intra-group feature compactness, individual-group center consistency, and inter-group feature separability to make features of the same group cluster in the embedding space and features of different groups move away; in the inference and matching phase, treating the feature sets of the query group and the candidate groups in the image library as unordered individual sets respectively, using an optimal subset matching strategy to measure the similarity between groups, and calculating the minimum matching distance between the query group and the candidate groups; sorting the candidate groups in the image library and outputting the result with the highest similarity to the query group; this invention improves accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer vision image retrieval technology, specifically to a cross-modal small-group pedestrian re-identification method based on optimal subset learning. Background Technology

[0002] Pedestrian re-identification technology has gradually expanded from single-person retrieval to the more practical re-identification of small groups of people. With the continuous development of deep learning and neural network architectures, small-group pedestrian re-identification has achieved remarkable results. However, most existing small-group pedestrian re-identification methods are based on visible light images from daytime, heavily relying on apparent features such as color and texture. Their performance deteriorates sharply in low-light environments such as nighttime, backlighting, or inclement weather. To achieve all-weather monitoring, practical systems often deploy dual-modal cameras (visible light and infrared). However, due to the fundamental differences in imaging principles and visual representation between visible light and infrared images, inconsistent cross-modal feature distributions result in existing models exhibiting weak generalization ability in cross-modal scenarios. Therefore, to overcome this limitation, recent research has gradually shifted its focus to visible light and infrared small-group pedestrian re-identification, fully utilizing the advantages of infrared images in low-light environments to provide a more reliable solution for small-group identification.

[0003] Cross-modal small-group pedestrian re-identification not only needs to address the inherent challenges of traditional small-group pedestrian re-identification tasks (such as changes in the number and layout of members within the group), but also must overcome the significant differences between two different sensor modalities. This modal difference stems from the different physical principles used in the acquisition process of visible light and infrared images. Visible light images are directly affected by natural lighting conditions, capturing the surface color and texture of objects, while infrared images are based on the thermal radiation of the target object, focusing more on the temperature distribution of the object. These two different physical properties lead to significant differences in pedestrian appearance, texture, brightness, and thermal distribution. Existing methods can generally be divided into two categories: (1) feature extraction methods based on deep learning, which aim to learn discriminative group embedding representations from group images and jointly model the member relationships, spatial layout, and relative positions within the group; (2) methods based on metric learning, which aim to learn a feature embedding space with strong generalization ability and explicitly optimize the relative distance relationship between features. However, these two types of methods cannot directly handle huge cross-modal differences, and their modeling of group structure is coarse-grained, failing to make full use of group-level semantics for metric optimization, resulting in difficulty in learning enough discriminative information. Summary of the Invention

[0004] Objective: This invention aims to provide a cross-modal small-group pedestrian re-identification method based on optimal subset learning. It improves and optimizes existing cross-modal small-group pedestrian re-identification algorithms by proposing an optimal subset matching strategy from a similarity modeling perspective. This strategy finds the optimal matching subset by explicitly searching for one-to-one correspondences between members of two groups, while globally penalizing unmatched individuals, thereby modeling group similarity and improving matching accuracy when the group structure changes dynamically. From a feature extraction perspective, a two-dimensional global refinement module is proposed. This module captures global dependencies and emphasizes modality-invariant discriminative identity information, thereby enhancing pedestrian features and improving the reliability of subsequent member-level matching. Furthermore, it improves and optimizes existing cross-modal small-group pedestrian re-identification algorithms by proposing a cross-modal intra-group consistency loss function from a feature optimization perspective. This enhances the intra-group compactness and inter-group separability of cross-modal features, resulting in more clearly structured and discriminative group features. Finally, it addresses the problems existing in the background technology.

[0005] Technical solution: The present invention provides a cross-modal small-group pedestrian re-identification method based on optimal subset learning, comprising the following steps:

[0006] (1) Construct a feature extraction network to receive a group to be identified containing visible light images and infrared images, and extract individual features of each pedestrian in the group through a backbone network with shared weights;

[0007] (2) In the feature extraction process of the backbone network, a two-dimensional global refinement mechanism is introduced. By parallel processing of spatial compression and global context modeling, the long-range dependencies of individual features in the spatial and channel dimensions are captured, and an enhanced set of individual features that is insensitive to modal differences is obtained.

[0008] (3) During the training and optimization phase of the network, cross-modal intra-group consistency constraints are used to guide the feature distribution. By combining intra-group feature density, individual and group center consistency, and inter-group feature separation, features of the same group are clustered in the embedding space and features of different groups are kept apart.

[0009] (4) In the reasoning and matching stage, the feature sets of the query group and the candidate group of the library are regarded as unordered individual sets respectively. The optimal subset matching strategy is used to measure the similarity between groups. The matching subset is determined by searching the one-to-one optimal correspondence between the members of the two sets. The remaining members in the case of mismatch are penalized based on global distance information. The minimum matching distance between the query group and the candidate group is calculated.

[0010] (5) Sort the candidate groups in the library according to the minimum matching distance and output the result with the highest similarity to the query group.

[0011] Further, step (2) is as follows: Construct a dual-branch structure. The first branch performs spatial downsampling and channel compression on the input feature map to generate query features. The second branch combines global pooling and channel projection to generate key features and value features. Use multi-head attention mechanism to aggregate global context information in the compressed space and generate an adaptive attention mask based on the aggregation result. Finally, reweight the original individual features element by element through residual connection.

[0012] Furthermore, step (2) also includes: after obtaining the global feature map output by the backbone network, the feature map is decomposed in a structured manner using a part-level spatial attention mechanism; this mechanism generates a set of two-dimensional weight maps corresponding to different semantic regions of the human body by learning, and applies each weight map to the global feature map to extract feature components focused on specific local parts; then, all extracted part-level feature components are spliced ​​with the undecomposed original global features to form a pedestrian individual representation that simultaneously covers overall contour information and local detail attributes.

[0013] Furthermore, a backbone network is constructed based on a convolutional neural network architecture, and the network parameters are fully shared across all levels for inputs in the visible light and infrared modes. The two-dimensional global refinement mechanism is embedded after multiple intermediate feature extraction stages of the backbone network in a mode-sharing manner.

[0014] Further, step (3) is as follows: intragroup tightness constraint is performed by maximizing the similarity between any two individual features within the same group; central consistency constraint is performed by minimizing the distance from each individual feature within the group to the center of the group features; intergroup separation constraint is performed by forcibly requiring that the average similarity between individuals within the same group maintains a distance from the average similarity between different groups.

[0015] Furthermore, in step (3), the gradient is calculated and the network parameters are updated using the identity label-based classification loss, the center clustering loss, and the cross-modal group intra-consistency constraint.

[0016] Further, step (4) is as follows: perform a linear space transformation on the distance matrix between individuals of the query group and the candidate group to solve for the one-to-one matching relationship that minimizes the total distance after the transformation, and determine the largest matchable common subset and its corresponding matching cost; for the remaining individual members that cannot be included in the common subset, calculate the penalty cost based on the number of remaining members and the maximum individual distance in the current matching scenario; add the matching cost of the common subset and the penalty cost to obtain the final matching distance between the two groups under the condition of changes in the number of members or layout.

[0017] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described herein.

[0018] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described herein.

[0019] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: It employs an optimal subset matching strategy to model group similarity. This strategy explicitly searches for the best one-to-one correspondence among group members while penalizing unmatched individuals, thereby achieving more accurate and finer-grained similarity estimation. To further improve the reliability of matching, a two-dimensional global refinement module is introduced into the feature extraction network to enhance member-level representations while suppressing noise caused by cross-modal differences. Furthermore, a cross-modal intra-group consistency loss is introduced in the feature optimization stage to enhance intra-group compactness and promote inter-group discriminability. This framework addresses the problem of insufficient matching robustness caused by changes in group structure in cross-modal scenarios, improving the accuracy of re-identification of small groups of pedestrians in this scenario. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this embodiment of the invention provides a cross-modal small-group pedestrian re-identification method based on optimal subset learning, including the following steps:

[0023] Step 1: Divide the dataset into training and testing sets. Pre-crop group images in the training set into individual images based on pedestrian labels. Randomly sample visible light and infrared individual images from the training set to form batch training data, and perform preprocessing operations such as resizing, horizontal flipping, and tensor normalization. Specifically: First, read the label information of group images and individual pedestrians from the training set. Crop the group images into individual images based on the pedestrian labels. Then, crop the individual images to 384x128 pixels and convert them into tensor data. Use random erasure, random horizontal flipping, and random cropping to avoid overfitting the network to the training data. Finally, standardize using the channel mean and standard deviation calculated from the ImageNet dataset, and then normalize the tensors to the range (0,1).

[0024] Step 2: Build a single-branch pedestrian re-identification network based on PyTorch and set the training parameters. Pair the training samples of the two modalities and use them alternately as the input of the network to extract pedestrian representations. In the single-person training data constructed in Step 1, a PK sampling strategy is used for random sampling. Each identity ID has the same number of visible light and infrared images. Pair them up across modalities and use them as the input of the backbone network to extract pedestrian features.

[0025] The framework of the cross-modal small-group pedestrian re-identification neural network consists of a backbone network, a two-dimensional global refinement module, and an optimal subset matching strategy. A ResNet50 model pre-trained on ImageNet was selected as the backbone network, and then fine-tuned on the CM-Group dataset. ResNet50 has five stages; the first stage serves as the input for visible light and infrared modal data, with both modalities sharing all five stages to form a single-branch pedestrian re-identification network. The two-dimensional global refinement module is integrated modally shared after the second and third stages of ResNet50. The optimal subset matching strategy is used during the later inference phase. The total training duration is 140 epochs, with an initial learning rate of 0.00035, which then decays by a factor of 0.1 at epochs 80 and 120. The inference phase, starting from epoch 120, is executed periodically every 5 epochs to select the optimal model. The testing phase is executed once after training is complete to evaluate the final model performance. A batch contains 128 images, of which 4 visible light images and 4 infrared images are randomly selected from 16 identities. We use Adam as the optimizer to optimize the model, with a weight decay of 0.0005.

[0026] Step 3: In the feature extraction stage, cross-modal pedestrian features are enhanced using two shared two-dimensional global refinement modules to obtain a more robust set of pedestrian features. These two-dimensional global refinement modules are inserted after the second and third stages of the ResNet50 backbone network to enhance individual features. By modeling global dependencies in individual appearance and emphasizing modality-invariant identity-discriminating cues through modality sharing, they alleviate modality differences and improve the reliability of subsequent member-level matching.

[0027] Given an individual's intermediate feature map (in , , (representing the length, width, and height of the feature map, respectively). The two-dimensional global refinement module first transforms the input features into three embeddings through a two-branch design: query... ,key Sum Specifically, the query branch is processed by a convolutional kernel with a stride of [missing value]. The convolutional projection performs spatial downsampling and channel compression, while and This is achieved through a global pooling operation and another convolutional projection (with a kernel size of 1). This compresses the input features into a compact latent space, preserving the most informative structural patterns while reducing computational cost. The resulting three embeddings are then flattened along the spatial dimensions, and a multi-head attention mechanism is used to aggregate globally relevant contextual information.

[0028]

[0029] in, This indicates a multi-head attention mechanism. This represents the output features of a multi-head attention mechanism. For each attention head, the query interacts with all key-value pairs to generate an attention map that jointly compresses both spatial and channel dimensions, measuring the correlation between features at different spatial locations for reweighted value embedding. This process allows each local region to capture complementary responses from other information-rich regions. Therefore, the network is able to model long-range dependencies across the entire feature map in the compressed space and enhance modality-sharing information.

[0030] The outputs of all attention heads are concatenated and projected back to the original channel dimension, then spatial upsampling is performed to restore the feature resolution. Given the reconstructed feature map, an attention mask is generated by applying the sigmoid activation function. :

[0031]

[0032] in and These represent linear mapping and spatial upsampling operations, respectively. This represents the sigmoid activation function. Finally, by analyzing the original input feature map... Perform element-wise reweighting to obtain the output features. :

[0033]

[0034] in This represents element-wise multiplication. The output features of the two-dimensional global refinement module are then used. Global features are obtained through feature extraction via the backbone network. Furthermore, it employs a location-level spatial attention mechanism to enable the model to focus on different local regions.

[0035] The part-level spatial attention mechanism first maps the output feature map to a 1x1 convolution. An attention mask (local area) The number is 6), then each mask is normalized and compared with the feature map. Element-wise multiplication is performed to obtain the features of the corresponding regions. Finally, all region-level features are concatenated with the global features to obtain the final feature vector. .

[0036]

[0037] in and represents 1x1 convolution and concatenation operations respectively, and p represents the local part index. The proposed two-dimensional global refinement module has two main functions: (1) By introducing global interactions in both spatial and channel dimensions, the network can capture long-range structural dependencies. (2) By generating an adaptive attention mask on the feature map, it emphasizes modality invariance and identity-related cues while suppressing the influence of noisy background regions. Therefore, the learned individual features become more discriminative, which helps with subsequent group matching.

[0038] Step 4: In the feature optimization stage, the model performs gradient backpropagation based on cross-modal intra-group consistency loss and baseline loss to optimize model parameters. First, the cross-modal intra-group consistency loss promotes intra-group compactness and inter-group segregation of cross-modal features, thereby generating more clearly structured and discriminative group representations, improving the accuracy of subsequent group matching. This loss works from three complementary perspectives: intra-group compactness, central consistency, and inter-group segregation.

[0039] Given a batch size of Feature embedding and their corresponding group tags ,make Indicates the first A set of individual characteristics for each group. This represents the group center. First, to encourage samples within the same group to be close to each other, we maximize the average pairwise similarity of samples within each group, resulting in an intragroup compactness loss. The calculation formula is expressed as:

[0040]

[0041] in, This represents the set of valid groups in a mini-batch. This represents the similarity between two feature embeddings, while and This represents different individuals within the same group. This term promotes feature aggregation among samples belonging to the same group. Secondly, to make the features of each group exhibit stronger internal compactness, we restrict the distribution range of group members around the corresponding group center, resulting in a center consistency loss. The calculation formula is expressed as:

[0042]

[0043] By minimizing the distance between an individual sample and its group center, this term suppresses the intra-group distribution range, thereby indirectly promoting the stabilization of feature distributions between groups. To prevent degradation collapse and maintain discriminability between different groups, we further impose a separation constraint between intra-group similarity and inter-group similarity, resulting in an inter-group separability loss. The calculation formula is expressed as:

[0044]

[0045] in, To find the maximum value function, and They represent the first The average within-group similarity and between-group similarity of the groups. This is a separation constraint value used to ensure that samples from the same group are more similar than samples from different groups. The final cross-modal intra-group consistency loss. Defined as:

[0046]

[0047] This loss aims to optimize the distribution in the feature space so that features within the same group are highly clustered while maintaining relative distance from features in different groups, thereby facilitating the subsequent matching process.

[0048] Secondly, the baseline loss includes cross-entropy loss and center clustering loss. The identity loss (cross-entropy loss) is used to constrain the network to extract pedestrian identity-related information, and its calculation process is as follows:

[0049]

[0050] in and These represent the number of visible light and infrared images in the current batch, respectively. and Representing pedestrian features in visible light and infrared modes, and These represent identity tags indicating corresponding characteristics, while and Is the same classifier used for pair? and The results of identity prediction. This represents the conditional probability of the true label being y given the output of the corresponding classifier. The center clustering loss is used to cluster features to their respective class centers, and its calculation process is as follows:

[0051]

[0052] in The label in the current batch is The mean of the characteristics, This is the number of identities in the current batch. It is the minimum interval between different types of centers.

[0053] Step 5: During the model inference phase, the minimum matching distance is calculated between the query group and the candidate group features in the image library. The optimal subset matching module is used to finely measure the similarity between groups, effectively improving the model's matching accuracy when group members are added or removed through fine-grained member matching, while eliminating the impact of changes in member layout. This module models each group as an unordered set of member features; therefore, when matching two groups subsequently, it is necessary to consider both the inconsistency in the number of members and the irrelevance of the member arrangement order. Assume the feature set of the query group Q is... The feature set of the image library group G is Let m and n represent the number of individuals in the two feature sets, respectively. The goal of the optimal subset matching module is to measure the similarity between the two groups by establishing the optimal correspondence between their members. Construct the distance matrices of Q and G. , where the matrix Each element Indicates the first Individual query and the first Distance between individual image library features.

[0054] when In this case, the matching problem can be considered a one-to-one assignment problem, with the goal of minimizing the matching cost. :

[0055]

[0056] in To perform the minimum value operation, This represents the operation of retrieving the value of the independent variable when the objective function is minimized. Represents a set of feasible correspondences between two groups, consisting of several pairs of indices. Composition, constraints Represents the matching scheme The number of matching pairs included, and stipulates that its number does not exceed the member size of any group; while This represents the optimal matching scheme, which is the scheme that maximizes the total matching distance while satisfying the constraints. The smallest matching set, This indicates the optimal matching scheme. The optimal subset matching distance between the two groups is obtained.

[0057] The above matching is achieved by performing multiple operations on the distance matrix. This is obtained by performing a linear transformation and extracting the corresponding group correspondences. Specifically, firstly, for... Perform row-by-row normalization (i.e., subtract the minimum value of each row), and then perform column-by-column normalization (i.e., subtract the minimum value of each column) to further adjust the matrix:

[0058]

[0059] in and These represent operations that take the minimum value of indices j and i, respectively. Let represent the row normalization coefficient of the i-th row of the distance matrix D, i.e., the minimum value of the i-th row. The column normalization coefficient of the j-th column of the distance matrix D is defined as the coefficient obtained by subtracting the minimum value of the corresponding row. Then, the minimum value among all elements in that column. After the above row-by-row and column-by-column normalization adjustments, the updated distance matrix is ​​obtained. Each of its elements is represented as After linear transformation Multiple zero-value elements will appear, each representing a candidate match with the minimum local cost.

[0060] based on The query individual characteristics are established by iteratively selecting zero-value elements. and individual characteristics of the image library The system establishes a correspondence between these pairs of matches, ensuring that each row and column can be selected at most once. When a conflict occurs (i.e., multiple zero values ​​appear in the same row or column), a complete and non-overlapping set of matches is prioritized. If the current... If the zero-value structure in the model is insufficient to construct a complete match, then a further linear transformation is performed. Adjustments are made by reintroducing additional zero-valued elements to generate feasible correspondences. This iterative process continues until a complete and non-overlapping set of matches is obtained, which is the optimal one-to-one match. The final matching cost is calculated as the sum of all distances under the optimal match.

[0061] when When the number of members in the two groups is unequal, a global dynamic penalty mechanism is introduced. Assuming the number of members in the query group (m) is greater than the number of members in the search group (n), the distance matrices of the query group Q and the image library group G are treated similarly. By performing multiple linear transformations and searching for "zero-cost" matches in the linear space, the feature set of the query group is obtained. subset of ( ) and image library group feature set Find the optimal one-to-one match and calculate the optimal subset matching distance. Subsequently, for the remaining set of members of the group to be queried... A penalty mechanism is introduced. Penalty items. Defined as the number of unmatched members Product with the maximum distance in the current matching scene:

[0062]

[0063] in This represents the distance metric function, while This represents the maximum feature distance among all cross-group feature pairs of Q and G, and this value will reflect the maximum possible feature difference between the two groups.

[0064] Final query - Gallery group distance Defined as the sum of the optimal subset matching distance and the penalty term:

[0065]

[0066] For each query group, obtain a list of distances to candidate groups, sort the list in ascending order, and return the top K most similar groups. For each candidate group feature in the graph library... The image library matching list is obtained by calculating the minimum matching cost between each image and the query group.

[0067]

[0068] in, This represents the image library matching list. This represents the optimal matching distance between the query group and the features of each candidate group in the image library, where t represents the total number of candidate groups in the image library. This list is sorted in ascending order, and the indices of the top K most similar groups are returned, forming the final re-identification output list. This process can be represented as:

[0069]

[0070] in This is a sorted index mapping function that returns a sequence of indices for a group. This is a list truncation operator, so the candidate groups in the output list are the top K image library groups that are most similar to the group to be queried.

Claims

1. A cross-modal small-group pedestrian re-identification method based on optimal subset learning, characterized in that, Includes the following steps: (1) Construct a feature extraction network to receive a group to be identified containing visible light images and infrared images, and extract individual features of each pedestrian in the group through a backbone network with shared weights; (2) In the feature extraction process of the backbone network, a two-dimensional global refinement mechanism is introduced. By parallel processing of spatial compression and global context modeling, the long-range dependencies of individual features in the spatial and channel dimensions are captured, and an enhanced set of individual features that is insensitive to modal differences is obtained. (3) During the training and optimization phase of the network, cross-modal intra-group consistency constraints are used to guide the feature distribution. By combining intra-group feature density, individual and group center consistency, and inter-group feature separation, features of the same group are clustered in the embedding space and features of different groups are kept apart. (4) In the reasoning and matching stage, the feature sets of the query group and the candidate group of the library are regarded as unordered individual sets respectively. The optimal subset matching strategy is used to measure the similarity between groups. The matching subset is determined by searching the one-to-one optimal correspondence between the members of the two sets. The remaining members in the case of mismatch are penalized based on global distance information. The minimum matching distance between the query group and the candidate group is calculated. (5) Sort the candidate groups in the library according to the minimum matching distance and output the result with the highest similarity to the query group.

2. The method for cross-modal small-group pedestrian re-identification based on optimal subset learning according to claim 1, characterized in that, Step (2) is as follows: Construct a dual-branch structure. The first branch performs spatial downsampling and channel compression on the input feature map to generate query features. The second branch combines global pooling and channel projection to generate key features and value features. A multi-head attention mechanism is used to aggregate global context information in a compressed space, and an adaptive attention mask is generated based on the aggregation result. Finally, the original individual features are reweighted element by element through residual connections.

3. The method for cross-modal small-group pedestrian re-identification based on optimal subset learning according to claim 1, characterized in that, Step (2) further includes: after obtaining the global feature map output by the backbone network, the feature map is decomposed into a structured form using a part-level spatial attention mechanism; the mechanism generates a set of two-dimensional weight maps corresponding to different semantic regions of the human body by learning, and applies each weight map to the global feature map to extract feature components focused on local parts; then, all extracted part-level feature components are concatenated with the undecomposed original global features to form a pedestrian individual representation that simultaneously covers overall contour information and local detail attributes.

4. The method for cross-modal small-group pedestrian re-identification based on optimal subset learning according to claim 1, characterized in that, The backbone network is constructed based on a convolutional neural network architecture. For inputs in visible light and infrared modes, the network parameters are fully shared across all levels. The two-dimensional global refinement mechanism is embedded in multiple intermediate feature extraction stages of the backbone network in a mode-sharing manner.

5. The method for cross-modal small-group pedestrian re-identification based on optimal subset learning according to claim 1, characterized in that, Step (3) is as follows: intragroup tightness constraint is performed by maximizing the similarity between any two individual features within the same group; central consistency constraint is performed by minimizing the distance from each individual feature within the group to the center of the group features; intergroup separation constraint is performed by forcibly requiring that the average similarity between individuals within the same group maintains a distance from the average similarity between different groups.

6. The method for cross-modal small-group pedestrian re-identification based on optimal subset learning according to claim 1, characterized in that, In step (3), the gradient is calculated and the network parameters are updated using the identity label-based classification loss, the center clustering loss, and the cross-modal group intra-consistency constraint.

7. The method for cross-modal small-group pedestrian re-identification based on optimal subset learning according to claim 1, characterized in that, Step (4) is as follows: Perform a linear space transformation on the distance matrix between individuals of the query group and the candidate group to solve the one-to-one matching relationship that minimizes the total distance after the transformation, and determine the largest matchable common subset and its corresponding matching cost. For the remaining individual members that cannot be included in the common subset, the penalty cost is calculated based on the number of remaining members and the maximum individual distance in the current matching scenario; the matching cost of the common subset and the penalty cost are added together to obtain the final matching distance between the two groups under the condition of changes in the number of members or layout.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images