A semantic isotropic enhancement method for generalizable pedestrian re-identification

Abstract

The application discloses a semantic isotropic enhancement method for generalizable pedestrian re-identification, and belongs to the technical field of computer vision. Multiple public pedestrian image datasets of different source domains are selected, and features are extracted by means of a backbone network and fine-grained information is retained. The channel mean and standard deviation of the features disturbed by Gaussian noise are used to reconstruct the features to complete style enhancement. A semantic distribution matrix is constructed for multiple groups of style features, the center distribution is calculated, and a distillation loss is constructed by using KL divergence to realize semantic distribution alignment. The mean and variance of the semantic gradient of multiple source domains are counted, and the semantic enhancement is completed by combining the randomly sampled translation gradient. Finally, a joint loss is constructed by fusing the identity classification loss and the distillation loss, and the model is obtained by iteratively training the network. The application can alleviate the field deviation and source domain overfitting problems, simultaneously enhance the style and semantic generalization ability, has small calculation burden, and can effectively improve the overall performance of pedestrian re-identification in an open scene.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and specifically to a semantic isotropic enhancement method for generalizable person re-identification. Background Technology

[0002] Pedestrian re-identification is a core computer vision technology in fields such as intelligent security, video surveillance, and smart cities. Its core objective is to retrieve and match the same pedestrian across multiple non-overlapping camera images. With the widespread deployment of public surveillance equipment, images from different cameras often exhibit significant differences in lighting, viewing angle, background, and shooting style. This domain shift problem has become a key bottleneck restricting the practical application of pedestrian re-identification models. Traditional supervised pedestrian re-identification models perform well only within a single training domain; when transferred to a completely new target domain that was not used in the training, recognition accuracy drops sharply. Domain generalization pedestrian re-identification technology has emerged to address this issue. This technology trains the model based on multiple labeled source domain data, requiring the model to adapt to unknown scenarios without target domain annotation information. It represents the mainstream approach to solving the cross-camera pedestrian matching problem. Currently, most mainstream solutions in the industry adopt approaches such as data augmentation, feature constraints, and network structure optimization to improve the generalization ability of models. However, most solutions focus too much on learning style-invariant features and ignore the diversity mining of core semantic information such as pedestrian posture, clothing, and actions. This can easily lead to problems such as pseudo-generalization and overfitting of source domain identity features, making it difficult to adapt to complex and ever-changing open-world application scenarios.

[0003] In recent years, numerous academic teams both domestically and internationally have conducted extensive research and produced a series of academic results on domain-generalized person re-identification. Most of these studies revolve around two main directions: data augmentation and feature optimization. The article "Domain-Generalized Person Re-identification Based on Knowledge Distillation and Mask Reconstruction," published in the *Journal of Sun Yat-sen University (Natural Science Edition)*, proposes a technical solution combining knowledge distillation and mask reconstruction. This solution mines cross-domain identity cues through a dual-branch network and enhances the model's cross-domain adaptability using a distillation strategy. This method achieves improved recognition accuracy on mainstream datasets. However, this solution focuses on constraining the overall image features and does not distinguish between style and semantic information, thus failing to effectively suppress the interference of low-frequency style information on high-frequency identity semantic features. The article "Domain-Generalized Person Re-identification Based on Dual-Style Enhancement," published in *Software Guide*, designs a style memory module and a dual-style loss function, and simulates... While diversifying style variations enhances a model's style generalization ability and effectively enriches the style dimensions of training data, it suffers from significant drawbacks. Its data augmentation strategy only operates at the image style level, failing to extend to pedestrian semantic features. When semantic changes occur in the target domain, such as pose or clothing, the model's generalization ability rapidly declines. Furthermore, a single style augmentation mode can lead to over-reliance on the source domain's style distribution, failing to address recognition errors caused by domain shifts at their semantic root. In addition, many international conference and journal papers proposing pose augmentation and generative data augmentation methods share the common flaw of prioritizing style over semantics, making it difficult to simultaneously improve both style and semantic diversity.

[0004] While academic research has progressed, a large number of patents related to domain-generalized person re-identification have emerged in China, proposing various solutions from an engineering implementation perspective. Among them, Chinese invention patent CN115661923B discloses a domain-generalized person re-identification method based on adaptive modeling of domain features. This solution constructs a backbone network, a sub-feature embedding network, and a domain-generalized feature fusion module. By adaptively combining features from different domains, it weakens inter-domain conflicts and improves the model's feature representation ability in multi-domain scenarios. This patented technology can effectively fuse multi-domain features and enhance the extraction effect of domain-invariant features, exhibiting good stability in conventional cross-domain scenarios. However, this technology does not introduce a targeted data augmentation mechanism and cannot actively expand the style and semantic diversity of the training data. When facing unfamiliar target domains with vastly different styles, the limitations of feature representation are amplified. Furthermore, this solution relies on network structure stacking to achieve domain generalization, resulting in a large number of model parameters and computational overhead. Deployment in scenarios with limited computing power, such as edge monitoring devices, is difficult. Moreover, it does not optimize semantic gradients, failing to guide the model to learn cross-domain generalized semantic features, and still suffers from insufficient semantic generalization ability.

[0005] While existing academic literature and patented technologies have optimized domain-generalized person re-identification models from the perspectives of feature extraction, data augmentation, and network structure, alleviating some domain bias issues, they have not completely solved common pain points in the industry. Existing technologies either simply enhance style while neglecting semantic diversity, or focus on domain feature fusion but lack efficient enhancement strategies. They generally suffer from drawbacks such as low-frequency style information interfering with identity semantic learning, models being prone to overfitting the source domain, and weak semantic generalization ability. At the same time, some solutions have complex network structures and high computational costs, making it difficult to balance recognition accuracy and deployment efficiency. Summary of the Invention

[0006] Based on the aforementioned technical problems, this application discloses a semantic isotropic enhancement method for generalizable pedestrian re-identification, comprising the following steps:

[0007] S1. Select multiple public pedestrian image datasets from different source domains, input the original pedestrian images into a preset backbone network for feature extraction, remove the last layer of spatial downsampling operation in the backbone network, and output the original pedestrian features corresponding to each pedestrian image.

[0008] S2. Perform style uncertainty enhancement processing on the original pedestrian features. Perturb the mean and standard deviation of each feature channel by Gaussian noise. Reconstruct the feature map based on the perturbated mean and standard deviation to generate multiple style-enhanced features with consistent semantic content but different styles.

[0009] S3. Construct a batch semantic distribution matrix for each style enhancement feature, obtain the central distribution by taking the mean of all semantic distribution matrices, construct a central distribution distillation loss based on KL divergence, and use the loss to constrain the semantic distribution of each style enhancement feature to align with the central distribution.

[0010] S4. Calculate the semantic gradients corresponding to the multi-source domain samples, and calculate the mean and variance of the semantic gradients of all source domains to determine the semantic variation range. Within this range, sample random quantities and shift the original semantic gradients to complete the semantic uncertainty enhancement and obtain the semantic enhancement features.

[0011] S5. Input the semantic enhancement features into the classifier, construct a joint loss function by combining the identity classification loss and the center distribution distillation loss, update the network parameters by backpropagation according to the joint loss, and obtain the trained domain generalized pedestrian re-identification model after the parameters converge.

[0012] Preferably, the original pedestrian features output by the backbone network in S2 are represented as a three-dimensional feature map. ;in This represents the total number of feature channels. The height of the feature map, The width of the feature map is given; the mean of all pixels in the spatial dimension of the feature map is calculated channel by channel, and the mean of the i-th pixel is given. The formula for calculating the mean of each channel is as follows: ,in This is the channel number currently being calculated; The coordinates of the feature map row, with a range of values. ; The coordinates of the feature map column, with a range of values. ; For the feature map at the th Channel, First line, number The feature pixel value at the column position.

[0013] Preferably, in S2, the standard deviation of all pixels in the feature map spatial dimension is calculated channel by channel based on the channel mean. The formula for calculating the standard deviation of each channel is as follows: ,in For the first The channel feature mean is used to characterize the dispersion of feature pixels within a single channel.

[0014] Preferably, in S2, Gaussian noise is introduced to perturb the mean and standard deviation of each channel to obtain the perturbed channel mean. Standard deviation of the channel after disturbance The formula is:

[0015]

[0016]

[0017] in: For the original number Channel mean For the original number Channel standard deviation; To act on the first The Gaussian noise sample value of the channel mean. To act on the first Gaussian noise sample value of the channel standard deviation; and All follow a normal distribution.

[0018] Preferably, in S2, the original feature map is reconstructed using the perturbed channel mean and channel standard deviation to obtain a single set of style-enhanced feature maps. The feature reconstruction calculation formula is as follows: ,in: This is the original 3D feature map; , They are the original number Channel mean, standard deviation; , These are the channel mean and channel standard deviation after the disturbance, respectively. The original features are standardized channel by channel; noise perturbation and feature reconstruction operations are repeated to generate multiple sets of style-enhanced features.

[0019] Preferably, in S3, the number of samples input to the network in a single run is set to the batch size. The original image was used to extract the batch of original feature sets. ; will the first Extract features from the style-enhanced images to obtain the corresponding batch style-enhanced feature set. , For style enhancement group number; calculate the number of the first group. In group style enhancement features, any sample within a batch With sample The semantic distribution weights are calculated using the following formula: ,in: Indicates the first Input data from each source domain; This is the Euclidean distance function, used to calculate the distance between two feature vectors; This is a temperature scaling parameter used to adjust the smoothness of the distribution. , The first Samples in group style enhancement features ,sample eigenvectors; For the first Samples under various styles Relative to the sample The semantic distribution weights.

[0020] Preferably, the total number of statistical style enhancements in S3 is ,right The semantic distribution weights corresponding to the style enhancement features are calculated by taking the arithmetic mean position by position to obtain the central distribution probability. The formula is: ,in: Increase the total number of style groups; Iterate through all style enhancement groups; For the first The semantic distribution weights corresponding to the group styles; This represents the baseline central distribution probability obtained after integrating all styles.

[0021] Preferably, in S3, the distribution matrix constructed using the central distribution probability is used as the soft label, and the KL divergence is used to measure the difference between a single style distribution and the central distribution to construct the central distribution distillation loss. The formula is: ,in: This is the KL divergence operator, used to measure the difference between two probability distributions; This is a baseline distribution generated based on the central distribution probability. For the first Semantic distribution corresponding to group styles; To enhance the overall quantity of styles, The batch size is used; triple summation iterates through style groups and all sample pairings within the batch.

[0022] Preferably, in S4, the total number of source domains participating in training is set to... , Indicates the first The semantic gradient directions corresponding to each source domain data point; first, calculate the mean of the semantic gradients of all source domain data. Then calculate the variance of the semantic gradient. The two formulas are as follows: , ,in The source domain index, with a value range of... ; This represents the total number of source domains. The overall mean of the semantic gradients of all source domains; The variance of the semantic gradient is used to characterize the range of semantic changes.

[0023] Within the scope of semantic variation, sample random variables that follow a standard Gaussian distribution. The enhanced semantic gradient is obtained by combining the gradient mean and gradient standard deviation with a shift: ,in: These are sampled values ​​from a standard Gaussian distribution. Let be the standard deviation corresponding to the semantic gradient variance; based on this formula, semantic gradient shift is performed to obtain semantically enhanced features.

[0024] Preferably, in S5, the semantically enhanced features are input into a nonparametric memory classifier, and a joint loss function is constructed by combining the pedestrian identity classification loss and the center distribution distillation loss. The joint loss formula is: ,in: Losses due to pedestrian identification classification The distillation loss weighting coefficient is preset by the user. The model employs a centrally distributed distillation loss; backpropagation is performed using a joint loss function to iteratively update all trainable parameters of the backbone network and classifier until the network parameters converge, thus completing model training.

[0025] Compared with the prior art, the technical solution of this application has the following technical effects:

[0026] This invention enhances style uncertainty by perturbing the mean and standard deviation of feature channels with Gaussian noise and reconstructing the features. This enriches the style representation of samples and effectively broadens the range of styles learned by the model. This method achieves style randomization based on feature statistics, which weakens the feature constraints brought by fixed styles, allowing the model to stably learn style features with universal attributes, and improving the overall adaptability of the model to styles in different scenarios.

[0027] This invention constructs semantic distributions corresponding to multiple sets of style features and calculates their central distributions. It then combines this with KL divergence to build a central distribution distillation loss, guiding the semantic distributions of various style features to align. This design eliminates semantic interference from different styles, focuses on the core semantic relationships of the samples themselves, standardizes feature learning directions, and enables the model to accurately capture essential semantic information unaffected by style.

[0028] This invention statistically analyzes the mean and variance of semantic gradients from multiple source domains, determines the range of semantic variation based on the statistical results, and shifts the semantic gradients to enhance semantic uncertainty. This processing method enriches the dimensions of semantic expression, expands the semantic boundaries perceived by the model, reduces the model's dependence on the inherent semantics of the source domain, and continuously strengthens the model's ability to mine cross-domain general semantic features.

[0029] This invention integrates identity classification loss and center distribution distillation loss to form a joint loss function, and iteratively updates network parameters based on the joint loss. The entire technical process is concise in structure and clear in computational logic. While achieving dual enhancement of style and semantics, it maintains low computational overhead, ensuring efficient operation of the model in various application scenarios and comprehensively improving the overall performance of person re-identification in open domains.

[0030] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0031] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0033] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0034] Figure 1 This is a schematic diagram of the overall branching process architecture of the semantic isotropic enhancement method for domain-generalized pedestrian re-identification according to the present invention.

[0035] Figure 2 This is a schematic diagram of the overall network operation structure of the multi-source domain parallel style enhancement, central distillation, and semantic gradient enhancement of the present invention;

[0036] Figure 3 This is a schematic diagram illustrating the principle of distilling and fusing multiple style enhancement features into a unified central semantic distribution according to the present invention.

[0037] Figure 4 This is a schematic diagram showing the change curves of recognition accuracy in the Market-1501 and DukeMTMC datasets under different styles of enhancement.

[0038] Figure 5 This is a schematic diagram showing the change curves of recognition accuracy on the Market-1501 and DukeMTMC datasets under different style-enhanced variances according to the present invention;

[0039] Figure 6 This is a schematic diagram of the two-dimensional feature distribution scatter points of the multi-source domain dataset after features are extracted by this method. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0041] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0042] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0043] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0044] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0045] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0046] Example 1

[0047] This embodiment mainly describes a semantic isotropic enhancement method for generalizable pedestrian re-identification, such as... Figure 1 As shown, it specifically includes:

[0048] S1. Select multiple public pedestrian image datasets from different source domains, input the original pedestrian images into a preset backbone network for feature extraction, remove the last layer of spatial downsampling operation in the backbone network, and output the original pedestrian features corresponding to each pedestrian image.

[0049] S2. Perform style uncertainty enhancement processing on the original pedestrian features. Perturb the mean and standard deviation of each feature channel by Gaussian noise. Reconstruct the feature map based on the perturbated mean and standard deviation to generate multiple style-enhanced features with consistent semantic content but different styles.

[0050] S3. Construct a batch semantic distribution matrix for each style enhancement feature, obtain the central distribution by taking the mean of all semantic distribution matrices, construct a central distribution distillation loss based on KL divergence, and use the loss to constrain the semantic distribution of each style enhancement feature to align with the central distribution.

[0051] S4. Calculate the semantic gradients corresponding to the multi-source domain samples, and calculate the mean and variance of the semantic gradients of all source domains to determine the semantic variation range. Within this range, sample random quantities and shift the original semantic gradients to complete the semantic uncertainty enhancement and obtain the semantic enhancement features.

[0052] S5. Input the semantic enhancement features into the classifier, construct a joint loss function by combining the identity classification loss and the center distribution distillation loss, update the network parameters by backpropagation according to the joint loss, and obtain the trained domain generalized pedestrian re-identification model after the parameters converge.

[0053] Furthermore, this application designs a semantic isotropic enhancement model (SIA) for open-domain person re-identification, such as... Figure 2 As shown, it is built on the ResNet-50 backbone and multi-source domain generalization training framework. To solve the problems of insufficient semantic generalization and easy overfitting of source domain identities in existing methods, the model integrates three core modules: style enhancement, semantic enhancement and distribution distillation, to achieve synergistic optimization of style invariance and semantic diversity.

[0054] Given a multi-source pedestrian image as input, features are first extracted using a backbone network. The last spatial downsampling layer is removed to preserve fine-grained feature information and enhance the richness of feature representation. Subsequently, a style enhancement module randomizes instance styles by perturbing the feature mean and variance with Gaussian random variables, breaking the limitations of source domain styles. Center distribution distillation loss (CeD-Loss) uses the center distribution of multi-style features as soft labels to distill the feature relationship distribution of each enhanced style, eliminating style interference and guiding the model to learn the direction of style-invariant gradients. Building upon this, semantic uncertainty enhancement (SeA) estimates the semantic distribution range using a Gaussian mixture model and shifts the training gradient along the effective semantic direction, avoiding overfitting of the model to source domain semantics. Finally, the enhanced features are input into a nonparametric memory classifier and trained using a joint loss function consisting of triplet loss and identity loss, outputting a pedestrian feature representation with strong cross-domain generalization ability, achieving accurate pedestrian matching in open-domain scenarios.

[0055] Furthermore, existing style enhancement methods eliminate spurious associations between features and labels by perturbing feature statistics, but ignore the semantic differences between the source and target domains. This can easily lead to the model overfitting to the source domain identity and make it difficult to achieve effective generalization in open domain scenarios. Therefore, an enhancement strategy that can discover effective semantic directions and improve semantic diversity is needed to guide the model to learn cross-domain general semantic features.

[0056] Inspired by the correlation between feature variance and semantic variation in generative models, SeA aims to enhance the semantic generalization ability of models by estimating the semantic distribution range and shifting the training gradient along the effective semantic direction. Its core idea is to capture semantic uncertainty by utilizing the semantic variance between source domain identities, and to generate diverse semantic shifts through random sampling, guiding the model to explore more meaningful semantic variations, thereby adapting to unseen target domains.

[0057] To obtain the effective range of semantic changes, SeA first calculates the semantic gradient direction statistics for all source domain data, including the mean and variance, to quantify semantic uncertainty. Formally, the estimated semantic range can be expressed as:

[0058]

[0059]

[0060] in, Let K be the semantic gradient direction of the j-th source domain data, and K be the total number of source domains used for training. It is the mean of the semantic gradients of all source domains. This is the variance of the semantic gradient, used to characterize the estimated range of semantic changes. The magnitude of the variance reflects the probability of changes in the direction of the corresponding semantic gradient, providing a basis for subsequent semantic augmentation. Although the underlying distribution of semantic shifts is unpredictable, the uncertainty estimate captured from the source domain identity can provide appropriate and meaningful data. To increase semantic diversity, SeA performs random sampling within the estimated semantic distribution range, shifting the original semantic gradient to generate the augmented semantic gradient. The specific formula is as follows:

[0061]

[0062] in, It follows a standard Gaussian distribution and is used to introduce random semantic offsets; The semantic variation range estimated earlier. Using this formula, the model can explore diverse semantic expressions along effective semantic directions, reducing dependence on source domain-specific semantics and thus improving generalization performance in open domain scenarios.

[0063] SeA does not require additional complex computational overhead; it achieves semantic enhancement solely through statistical analysis of semantic gradients and random sampling. Its advantages lie in its ability to accurately capture the range of semantic variations. The statistical characteristics of the source domain data ensure the effectiveness of semantic enhancement. Furthermore, it can generate diverse semantic expressions through random offsets, effectively mitigating overfitting. Complementing style enhancement methods, it strengthens semantic generalization while eliminating style interference, providing a more reliable feature learning direction for open-world person re-identification.

[0064] Furthermore, exploring the style-invariant gradient direction can improve the model's style generalization ability, which is also the basis for semantic uncertainty enhancement. To address the problem of limited style diversity in the source domain, this paper enriches the training data by changing the style of instances. As shown in style transfer research, changes in feature statistics can encode style information. Therefore, this paper achieves style enhancement by perturbing feature statistics in uncertain directions while preserving semantic content.

[0065] Instance normalization (IN) is widely used in image style removal in transfer models. Its core principle is to transform data values ​​with different distributions and scales into a form with uniform numerical properties according to specific rules. This effectively regularizes feature representations in heterogeneous domains, giving the trained model stronger domain adaptability and better suitability for unseen new domains. A typical example is batch normalization (BN), which uses statistical features calculated from mini-batch samples to normalize the input features. This can be categorized into group normalization (GN) (grouping mini-batch samples) and IN (performing a BN-like calculation on a single sample). IN better captures the features of each individual sample instance, making it more suitable for extracting domain-invariant features. Given a feature map... Where C, H, and W represent the number of channels, height, and width, respectively. This paper denotes the mean and standard deviation of the channel dimension features as follows: , The calculation formula is as follows:

[0066]

[0067]

[0068] in As many style transfer studies have pointed out, the mean of instances and standard deviation To encode style information and enrich the source domain style, this paper perturbs the mean value of each instance space dimension with Gaussian noise. and variance This enables style randomization.

[0069] Specifically, during training, the mean and variance are perturbed by Gaussian noise to randomize the style:

[0070]

[0071]

[0072] in and Sampled from a normal distribution:

[0073]

[0074]

[0075] In the formula , , and , These are the mean and variance of the normal distribution, respectively.

[0076] Using the enhanced style restoration example, the formula is: ,in This indicates an image with enhanced style.

[0077] Furthermore, traditional person re-identification (Re-ID) loss functions primarily focus on extracting discriminative information, but struggle to leverage style enhancement to uncover style-invariant gradient directions. After style enhancement, the feature relationships between samples of different styles differ, making them susceptible to style interference and prone to deviating from the core semantics. This hinders the model from learning cross-style universal feature representations, impacting open-domain generalization performance. Therefore, a distillation mechanism is needed to align the feature relationship distributions of different enhancement styles, eliminate interference from style differences, and guide the model to explore style-invariant gradient directions.

[0078] This paper constructs a central distribution as a soft label and distills the feature relation distribution of each augmentation style to align style with semantic relationships. Specifically, it first extracts features from the original data and multiple augmentation style data, calculates the feature-level distribution matrix for each augmentation style to capture the semantic associations between samples, and then averages the distribution matrices of all augmentation styles to generate a central distribution. This central distribution is not affected by a single style and can reflect the core semantic relationships. Then, the central distribution is used as a soft label, and the distillation loss guides the feature distribution of each augmentation style to align with the central distribution, forcing the model to ignore style differences and focus on the semantic essence. Ultimately, this helps the model discover style-invariant gradient directions and improve cross-domain generalization ability. Figure 3 As shown.

[0079] Assume the original data The extracted features are Where B is the batch size, and the first... One enhanced style data The extracted features are The total number of all enhanced styles is Furthermore, the augmented data and the original data have the same semantic content. Therefore, the semantic distribution can be represented as:

[0080]

[0081] in Indicates the first The semantic distribution matrix of the augmented data, and , Represents the source domain Data;

[0082] The expression is as follows: ,in, Represents the source domain Data, For Euclidean distance, This is a temperature scaling parameter used to adjust the smoothness of the distribution. Indicates enhanced style The feature relationship weights between sample a and sample b are determined. Then, the distribution matrices of all enhancement styles are averaged to obtain the central distribution probability: , where n is the number of enhanced styles.

[0083] The center distribution integrates semantic relations from all enhanced styles, reflecting stable core associations between samples and unaffected by single-style noise. Soft labels are generated by calculating the predicted distribution of the center value.

[0084]

[0085] in This represents the temperature scaling parameter. Subsequently, this paper uses the center distribution matrix as a soft label, and applies it to the proposed center distribution distillation loss. Distillation is performed on each enhancement style, and the loss expression is as follows:

[0086]

[0087] In the formula Indicates the first A semantic distribution of enhanced style, Denotes KL divergence, Represents the source domain The data.

[0088] The semantically enhanced features are input into a nonparametric memory classifier, and a joint loss function is constructed by combining pedestrian identity classification loss and center distribution distillation loss. The joint loss formula is: ,in: Losses due to pedestrian identification classification The distillation loss weighting coefficient is preset by the user. The model employs a centrally distributed distillation loss; backpropagation is performed using a joint loss function to iteratively update all trainable parameters of the backbone network and classifier until the network parameters converge, thus completing model training.

[0089] This embodiment achieves style enhancement by perturbing feature statistics with Gaussian noise, effectively enriching the style dimensions of the samples. Simultaneously, it utilizes center distribution distillation to complete multi-style semantic alignment, eliminating semantic interference caused by style. Semantic enhancement is achieved by combining semantic gradient statistics and gradient shifting, expanding the semantic expression range and reducing the model's dependence on source domain semantics. The entire scheme combines multiple loss functions for network training, resulting in low overall computational overhead. It simultaneously improves the model's style adaptability and semantic expression capabilities, significantly enhancing domain generalization performance and improving the overall performance of person re-identification in open scenarios.

[0090] Example 2, based on Example 1, details the technical evaluation of this application. Extensive experiments were conducted on four large-scale datasets: Market-1501, DukeMTMC-reID, CUHK03, and MSMT17. These datasets are widely used in people re-identification tasks. All datasets contain over 1000 identities and thousands of images. The Market-1501 dataset contains 32668 pedestrian images of 1501 identities, captured by 6 different cameras. The DukeMTMC-reID dataset contains 36411 images of 1812 identities. The CUHK03 dataset contains 14097 images of 1467 identities, captured by 8 and 6 high-resolution cameras respectively. The MSMT17 dataset contains 126441 images of 4101 identities, captured by 15 different cameras, making it the largest people re-identification dataset to date. Following the approach described in the literature, this paper uses a leave-one-domain-out protocol to divide the four datasets into training and testing domains. Specifically, the four datasets are divided into two parts: three datasets are used as the source training domain, and the remaining one is used as the open target domain.

[0091] This paper employs two evaluation protocols to quantitatively assess model performance. Protocol 1 uses only the training set from the source domain for model training, while Protocol 2 uses both the training and test sets from the source domain. During the testing phase, this paper follows the standard evaluation protocol for the test dataset, using mean Average Precision (mAP) and Cumulative Matching Characteristic (CMC) curves as quantitative evaluation metrics. All datasets utilize their original training-test split protocols for model performance evaluation; however, the CUHK03 dataset employs a novel training-test split protocol proposed in the literature, which includes both labeled and detection versions. This paper selects the more challenging detection version for experimentation.

[0092] This paper selects ResNet-50 as the backbone network and initializes it based on an ImageNet pre-trained model. To enrich the granularity of the final feature maps, this paper follows the approach in the literature and removes the final spatial downsampling operation in the backbone network. The input images are resized to 256×128 and data augmentation is performed through random flipping and random cropping. During training, the batch size is set to 64, containing 16 identities, with 4 images corresponding to each identity; the boundary value of the triplet loss is set to 0.3 in the experiment. For the identity loss function in the memory database, the update rate is... Set to 0.2, temperature factor Set to 0.05; the amount of style enhancement. Set it to 4.

[0093] The comparison is performed under Protocol-1. To verify the superiority of the proposed method, this paper compares it with several existing state-of-the-art methods on four widely used benchmark datasets, and the results are shown in Table 1. Here, M, D, C, and MS represent the Market-1501, DukeMTMC-ReID, CUHK03, and MSMT17 datasets, respectively. Some recent methods aim to learn generalization models from a single domain; this paper reproduces these methods under a multi-source DG ReID setting for comparison. The results for DSU and META are reproduced by running their publicly available code. It should be noted that no post-processing techniques such as re-ranking are used in any of the reported results.

[0094] Comparison on the Market-1501 dataset. In the Market-1501 dataset test experiments, the source domain data used were from three datasets: DukeMTMC, CUHK03, and MSMT17. The experimental setup is labeled in Table 1 as "MS+D+C→M". Under this experimental setup, the proposed SIA method was compared with several advanced algorithms in the field, covering mainstream methods such as QAConv, SNR, M³L, DSU, META, and the OSNet series. Experimental results show that the proposed SIA method performs best under this setup: the model based on the ResNet50 base network achieves a rank-1 accuracy of 81.2% and a mAP of 56.8%; while the model based on the ResNet50-IBN base network achieves a rank-1 accuracy of 81.7% and a mAP of 56.9%. Among the comparative methods, the META method based on the ResNet50 base network achieved the best performance, with a rank-1 accuracy of 78.1% and an mAP of 53.2%. In contrast, the SIA method in this application outperforms the META method by 3.1% in rank-1 accuracy and improves the mAP by 3.6%. Similarly, the meta-learning method M³L, also based on the ResNet50 base network, achieved a rank-1 accuracy of 76.5% and an mAP of 51.1%, demonstrating a significant performance advantage for the SIA method. Even though the M³L method improved its performance by adopting the ResNet50-IBN base network, achieving a rank-1 accuracy of 78.3% and an mAP of 52.5%, the SIA method in this application still outperformed it by 3.4% and 4.4 points in these two metrics, respectively. In summary, the experimental results on the Market-1501 dataset fully demonstrate that the SIA method proposed in this application performs outstandingly in the domain-generalized person re-identification task, significantly outperforming other comparative methods.

[0095] Table 1 shows the model results on Market-1501 (M), DUkeMTMC (D), CUHK03 (C), and MSMT17 (MS).

[0096] Comparison on the DukeMTMC dataset. When the DukeMTMC dataset was used as the test set, the source domain data consisted of three datasets: Market-1501, CUHK03, and MSMT17. This experimental setup is marked as "MS+M+C→D" in Table 1. Experimental results show that the SIA method of this application still maintains the best performance on this dataset: based on the ResNet50 base network, the rank-1 accuracy is 70.2%, and the mAP is 52.1%; based on the ResNet50-IBN base network, the rank-1 accuracy improves to 70.8%, and the mAP reaches 52.6%. Among the comparison methods, the META method based on the ResNet50 base network performs better, with a rank-1 accuracy of 68.7% and an mAP of 49.7%. Compared with this, the SIA method of this application improves the rank-1 accuracy by 1.5% and the mAP by 2.4%. On this dataset, the M³L method achieves similar performance based on ResNet50 and ResNet50-IBN, with rank-1 accuracies of 67.1% and 67.2%, and mAP of 48.2% and 48.8%, respectively. Compared to the best results achieved by the M³L method, the SIA method in this application achieves a rank-1 accuracies 3.6% higher and an mAP accuracies 3.8% higher. Furthermore, the SIA method also demonstrates greater performance advantages on other datasets, validating the stronger domain generalization capability of the proposed method.

[0097] Furthermore, under the Protocol-2 experimental setup, this paper specifically compared the SIA method with the RaMoE method, and the experimental results are shown in Table 2. As the data in the table shows, the performance advantage of the SIA method becomes more pronounced with the increase in the number of training samples: on the Market1501 dataset, the SIA method achieves an mAP of 61.4% and a rank-1 accuracy of 85.2%; on the MSMT17 dataset, the mAP is 20.7% and the rank-1 accuracy is 45.9%. Compared with the RaMoE method, the SIA method improves the mAP by 4.9% and 7.2% respectively on the above two datasets, and the rank-1 accuracy by 3.2% and 11.8% respectively.

[0098] Comparison on the MSMT17 dataset. When the test set is the MSMT17 dataset, the source domain data consists of three datasets: Market-1501, CUHK03, and DukeMTMC. This experimental setup is denoted as "M+D+C→MS". The MSMT17 dataset is one of the largest and most challenging datasets in the field of person re-identification. The SIA method in this application still achieved satisfactory performance under this experimental setup: based on the ResNet50 base network, the rank-1 accuracy was 42.3%, and the mAP was 18.4%; based on the ResNet50-IBN base network, the rank-1 accuracy improved to 43.4%, and the mAP reached 19.5%. Among the comparison methods, the META method based on the ResNet50 base network still performed well, with a rank-1 accuracy of 37.8% and an mAP of 15.7%. Compared with this, the SIA method in this application improved the rank-1 accuracy by 4.5% and the mAP by 2.7%. The meta-learning method M³L, based on the ResNet50 base network, has relatively weak performance, with a rank-1 accuracy of only 32.0% and an mAP of 13.1%, which are 10.3% and 5.3% lower than the ResNet50 version of the SIA method, respectively. Even though the M³L method improves its performance by using the ResNet50-IBN base network, achieving a rank-1 accuracy of 37.1% and an mAP of 15.4%, the ResNet50-IBN version of the SIA method in this application still outperforms it by 6.3% and 4.1% in these two metrics, respectively.

[0099] It is worth noting that the OSNet method, with the help of instance normalization (IN) technology, achieved a rank-1 accuracy of 40.2% on this dataset, which is one of the better performing methods among all comparison methods. However, it still has a significant performance gap compared with the SIA method proposed in this application. Meanwhile, the SIA method shows a more significant advantage on other datasets, further validating the performance stability of the method proposed in this application.

[0100] Comparison on the CUHK03 dataset. When conducting test experiments on the CUHK03 dataset, the source domain data came from three datasets: Market-1501, DukeMTMC, and MSMT17. This experimental setup is marked as "MS+M+D→C" in Table 1. Experimental results show that the SIA method of this application also exhibits significant performance superiority under this setup: based on the ResNet50 base network, the rank-1 accuracy is 35.2%, and the mAP is 33.4%; based on the ResNet50-IBN base network, the rank-1 accuracy improves to 35.8%, and the mAP reaches 33.7%. Among the comparative methods, the META method achieved the best performance under this experimental setup, with a rank-1 accuracy of 32.5% and an mAP of 31.8%; compared to this, the ResNet50 version of the SIA method in this application shows a 2.7% improvement in rank-1 accuracy and a 1.6% improvement in mAP. The meta-learning method M³L achieves similar performance based on ResNet50 and ResNet50-IBN, with rank-1 accuracy of 31.9% and 31.6% respectively, and mAP of 30.9% and 31.4% respectively. The ResNet50-IBN version of the SIA method in this application achieves a rank-1 accuracy 4.2% higher and an mAP accuracy 2.3% higher than the best results of the M³L method.

[0101] To further verify the generalization ability of the SIA method in this application to unseen target domains, this paper supplements the comparison between the SIA method and all the comparison methods in the table, on the average performance metrics of the four datasets mentioned above. Experimental results show that the average performance of the SIA method in this application is also at the leading level on the four datasets: when based on the ResNet50 base network, the average rank-1 accuracy is 57.2%, and the average mAP is 40.2%; when based on the ResNet50-IBN base network, the average rank-1 accuracy is improved to 57.9%, and the average mAP reaches 40.7%. Compared with META, the SIA method based on the ResNet50-IBN base network has an average rank-1 metric and an average mAP metric that are 3.6% and 3.1% higher, respectively. Furthermore, the meta-learning method M³L, based on ResNet50 and ResNet50-IBN networks, achieved average rank-1 accuracies of 51.9% and 53.6%, respectively, and average mAP of 35.8% and 37.0%, respectively. The ResNet50-IBN version of the SIA method in this application, compared to the best average performance of the M³L method, shows an average rank-1 accuracy 4.3% higher and an average mAP accuracy 3.7% higher. These experimental results fully demonstrate that the SIA method proposed in this application can effectively generalize to any unseen target domain and possesses excellent domain generalization performance.

[0102] Table 2 shows the comparative experiments under Protocol 2.

[0103] The comparison was conducted under Protocol-2. In the second experimental setup (Protocol-2), both the training and test sets from the source domain were used to train the model simultaneously, thus involving more training data and allowing for more thorough verification of the model's generalization performance. Under this experimental setup, the proposed SIA method was primarily compared with RaMoE, and the results are shown in Table 2. The proposed SIA method still exhibits superior performance under this experimental setup, fully demonstrating the effectiveness and generalization ability of the method. On the Market-1501 dataset, the SIA method based on ResNet50-IBN achieved a rank-1 accuracy of 85.2% and a mAP of 61.4%, which is 3.2% and 4.9% higher than RaMoE's 82.0% rank-1 and 56.5% mAP, respectively, representing a significant performance breakthrough. On the MSMT17 dataset, which is currently the largest dataset with the most significant domain shifts, the SIA method based on ResNet50-IBN achieved a rank-1 accuracy of 45.9% and a mAP of 20.7%, far exceeding RaMoE's 34.1% rank-1 and 13.5% mAP. The rank-1 accuracy was improved by 11.8%, and the mAP was improved by 7.2%, demonstrating strong adaptability in open domain scenarios with large domain gaps. While RaMoE achieved results similar to the SIA method in this application on the DukeMTMC-reID and CUHK03 datasets, the SIA method in this application significantly outperformed the two key datasets Market-1501 and MSMT17, fully demonstrating that the proposed method has stronger domain generalization ability in open-world pedestrian re-identification tasks, can better adapt to open target domains with different distributions, and effectively alleviates the performance degradation problem caused by domain shift.

[0104] This application conducts ablation experiments based on the ResNet50 backbone network to analyze the effectiveness of each component of the proposed method. The ablation experiments were conducted under two settings: "MS+D+C→M" and "MS+M+C→D". The experimental results for different components are shown in Table 3.

[0105] (1) Analysis of the effectiveness of style enhancement: When based on the ResNet50 backbone network, the performance is improved to 77.9% rank-1 and 52.4% mAP, respectively, achieving significant improvements of 3.1% rank-1 and 4.4% mAP and 3.1% rank-1 and 4.8% mAP. This result shows that style enhancement technology can effectively enrich the style of the source domain and help the model to discover style-invariant gradient directions.

[0106] (2) Analysis of the effectiveness of central distribution distillation. This application proposes the central distribution distillation loss, which aims to obtain the style-invariant gradient direction by using style enhancement. As shown in Table 3, this application combines style enhancement (StA) with central distribution distillation loss to learn style-invariant features in a collaborative manner, thereby further improving the generalization ability of the model. When based on the ResNet50 backbone network, under the setting of "MS+D+C→M", adding central distribution distillation loss improves the rank-1 of StA by 2.0%, and under the setting of "MS+M+C→D", the rank-1 is improved by 1.2%. This result verifies that the proposed central distribution distillation loss can effectively improve the style generalization ability of the model.

[0107] Table 3 Ablation Studies of Each Module

[0108] (3) Effectiveness analysis of semantic uncertainty augmentation: The proposed semantic uncertainty augmentation (SeA) technique aims to adjust the training gradient along the effective semantic direction to increase semantic diversity. As shown in Table 3, after adding SeA, the rank-1 of the model under the two settings of "MS+D+C→M" and "MS+M+C→D" is further improved to 79.1% and 68.1%, respectively, achieving a first-order matching rate improvement of 1.8% and 0.9%. Similar performance improvement can also be achieved when based on the ResNet50 backbone network. This result shows that SeA helps the model learn generalized semantics and avoids the model overfitting the source domain identity. After combining style enhancement, center distribution distillation loss and semantic uncertainty augmentation, the model performance is further improved. On the Market1501 dataset, the rank-1 reaches 81.2% and the mAP reaches 56.8%; on the DukeMTMC dataset, the rank-1 reaches 70.2% and the mAP reaches 52.1%.

[0109] This application conducted a sensitivity study on the variance of Gaussian noise and the amount of style enhancement in style enhancement, and the results are as follows: Figure 4 and Figure 5 As shown in the figure, it can be seen that as the variance of Gaussian noise in style enhancement increases, the model performance first shows an upward trend and then decreases, as... Figure 4 As shown, Figure 4 (a) and Figure 4 (b) Validation on Market 1501 and DukeMTMC respectively shows that the best performance is achieved when the variance is set to 0.1; the number of style enhancements also shows a similar trend, such as Figure 5 As shown, Figure 5 (a) and Figure 5 (b) Validations on Market 1501 and DukeMTMC show that the model performance is optimal when the number of style enhancements is set to 4. This indicates that increasing the number of style enhancements can effectively alleviate the limitations of source domain styles and improve the model's style generalization ability. To verify the effectiveness of the proposed method, this application uses the T-SNE data dimensionality reduction method to conduct feature visualization analysis on four benchmark datasets, such as... Figure 6 As shown, specifically, image samples are randomly selected from each dataset, and after forward propagation and pooling operations in the backbone network, deep features for pedestrian identification are obtained. These high-dimensional features are then reduced to a two-dimensional space, and feature points from the same dataset are labeled with the same color. The visualization results are shown in the figure. The experimental model was trained on three source domain datasets: Market-1501, DukeMTMC-reID, and MSMT17. The visualization results are shown in the figure. The experimental model was trained on three source domain datasets: Market-1501, DukeMTMC-reID, and MSMT17, and tested on the CUHK03 target domain.

[0110] As can be seen from the visualization results, the test set samples in the two-dimensional feature space are distributed within the range of the three training sets, indicating that the model trained on the multi-source domain can fit the data distribution of the target domain to a certain extent, thus possessing corresponding generalization performance for unknown domains. Meanwhile, the features of the four datasets are intertwined in space, making them difficult to effectively distinguish using simple linear or nonlinear boundaries.

[0111] This phenomenon reflects that the proposed method can learn feature representations with a certain degree of domain invariance, mapping samples from different data distributions to relatively concentrated feature regions. Furthermore, the three source domain datasets still exhibit a certain degree of discrete distribution on the two-dimensional plane, reflecting inherent inter-domain differences. This phenomenon indirectly confirms the high challenge and complexity of the domain-generalized person re-identification task, while also revealing its significant research value and broad application prospects in complex open scenarios.

[0112] This application proposes a semantic isotropic enhancement method for open-world person re-identification. It introduces semantic uncertainty enhancement, utilizes posterior probability to identify meaningful gradient directions, and guides the training gradient to adjust along effective semantic directions to improve model generalization ability. By randomizing instance styles using Gaussian random variables, arbitrary style transfer is achieved, thus overcoming the limitations of source domain styles. A center distillation loss is proposed to mine style-invariant gradient directions, providing support for improved generalization ability and semantic enhancement. Extensive experiments validate the effectiveness of this framework in training generalized person re-identification models.

[0113] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A semantic isotropic enhancement method for generalizable person re-identification, characterized in that, Includes the following steps: S1. Select multiple public pedestrian image datasets from different source domains, input the original pedestrian images into a preset backbone network for feature extraction, remove the last layer of spatial downsampling operation in the backbone network, and output the original pedestrian features corresponding to each pedestrian image. S2. Perform style uncertainty enhancement processing on the original pedestrian features. Perturb the mean and standard deviation of each feature channel by Gaussian noise. Reconstruct the feature map based on the perturbated mean and standard deviation to generate multiple style-enhanced features with consistent semantic content but different styles. S3. Construct a batch semantic distribution matrix for each style enhancement feature, obtain the central distribution by taking the mean of all semantic distribution matrices, construct a central distribution distillation loss based on KL divergence, and use the loss to constrain the semantic distribution of each style enhancement feature to align with the central distribution. S4. Calculate the semantic gradients corresponding to the multi-source domain samples, and calculate the mean and variance of the semantic gradients of all source domains to determine the semantic variation range. Within this range, sample random quantities and shift the original semantic gradients to complete the semantic uncertainty enhancement and obtain the semantic enhancement features. S5. Input the semantic enhancement features into the classifier, construct a joint loss function by combining the identity classification loss and the center distribution distillation loss, update the network parameters by backpropagation according to the joint loss, and obtain the trained domain generalized pedestrian re-identification model after the parameters converge.

2. The semantic isotropic enhancement method for generalizable person re-identification according to claim 1, characterized in that, The original pedestrian features output by the backbone network in S2 are represented as a three-dimensional feature map. ;in This represents the total number of feature channels. The height of the feature map, The width of the feature map is given; the mean of all pixels in the spatial dimension of the feature map is calculated channel by channel, and the mean of the i-th pixel is given. The formula for calculating the mean of each channel is as follows: ,in This is the channel number currently being calculated; The coordinates of the feature map row, with a range of values. ; The coordinates of the feature map column, with a range of values. ; For the feature map at the th Channel, First line, number The feature pixel value at the column position.

3. The semantic isotropic enhancement method for generalizable pedestrian re-identification according to claim 2, characterized in that, In S2, the standard deviation of all pixels in the feature map spatial dimension is calculated channel by channel based on the channel mean. The formula for calculating the standard deviation of each channel is as follows: ,in For the first The channel feature mean is used to characterize the dispersion of feature pixels within a single channel.

4. The semantic isotropic enhancement method for generalizable person re-identification according to claim 3, characterized in that, In S2, Gaussian noise is introduced to perturb the mean and standard deviation of each channel to obtain the perturbed channel mean. Standard deviation of the channel after disturbance The formula is: in: For the original number Channel mean For the original number Channel standard deviation; To act on the first The Gaussian noise sample value of the channel mean. To act on the first Gaussian noise sample value of the channel standard deviation; and All follow a normal distribution.

5. The semantic isotropic enhancement method for generalizable pedestrian re-identification according to claim 4, characterized in that, In S2, the original feature map is reconstructed using the perturbed channel mean and channel standard deviation to obtain a single set of style-enhanced feature maps. The feature reconstruction calculation formula is as follows: ,in: This is the original 3D feature map; , They are the original number Channel mean, standard deviation; , These are the channel mean and channel standard deviation after the disturbance, respectively. The original features are standardized channel by channel; noise perturbation and feature reconstruction operations are repeated to generate multiple sets of style-enhanced features.

6. The semantic isotropic enhancement method for generalizable person re-identification according to claim 5, characterized in that, In S3, the number of samples input to the network in a single run is set to the batch size. The original image was used to extract the batch of original feature sets. ; will the first Extract features from the style-enhanced images to obtain the corresponding batch style-enhanced feature set. , For style enhancement group number; calculate the number of the first group. In group style enhancement features, any sample within a batch With sample The semantic distribution weights are calculated using the following formula: ,in: Indicates the first Input data from each source domain; This is the Euclidean distance function, used to calculate the distance between two feature vectors; This is a temperature scaling parameter used to adjust the smoothness of the distribution. , The first Samples in group style enhancement features ,sample eigenvectors; For the first Samples under various styles Relative to the sample The semantic distribution weights.

7. The semantic isotropic enhancement method for generalizable person re-identification according to claim 6, characterized in that, The total number of statistical style enhancements in S3 is ,right The semantic distribution weights corresponding to the style enhancement features are calculated by taking the arithmetic mean position by position to obtain the central distribution probability. The formula is: ,in: Increase the total number of style groups; Iterate through all style enhancement groups; For the first The semantic distribution weights corresponding to the group styles; This represents the baseline central distribution probability obtained after integrating all styles.

8. The semantic isotropic enhancement method for generalizable person re-identification according to claim 7, characterized in that, In S3, the distribution matrix constructed from the central distribution probability is used as the soft label. KL divergence is used to measure the difference between a single style distribution and the central distribution, and a central distribution distillation loss is constructed. The formula is: ,in: This is the KL divergence operator, used to measure the difference between two probability distributions; This is a baseline distribution generated based on the central distribution probability. For the first Semantic distribution corresponding to group styles; To enhance the overall quantity of styles, The batch size is used; triple summation iterates through style groups and all sample pairings within the batch.

9. A semantic isotropic enhancement method for generalizable person re-identification according to claim 8, characterized in that, In S4, the total number of source domains participating in training is set to... , Indicates the first The semantic gradient directions corresponding to each source domain data point; first, calculate the mean of the semantic gradients of all source domain data. Then calculate the variance of the semantic gradient. The two formulas are as follows: , ,in The source domain index, with a value range of... ; This represents the total number of source domains. The overall mean of the semantic gradients of all source domains; The variance of the semantic gradient is used to characterize the range of semantic changes. Within the scope of semantic variation, sample random variables that follow a standard Gaussian distribution. The enhanced semantic gradient is obtained by combining the gradient mean and gradient standard deviation with a shift: ,in: These are sampled values ​​from a standard Gaussian distribution. Let be the standard deviation corresponding to the semantic gradient variance; based on this formula, semantic gradient shift is performed to obtain semantically enhanced features.

10. A semantic isotropic enhancement method for generalizable person re-identification according to claim 9, characterized in that, In S5, semantically enhanced features are input into a nonparametric memory classifier, and a joint loss function is constructed by combining pedestrian identity classification loss and center distribution distillation loss. The joint loss formula is: ,in: Losses due to pedestrian identification classification The distillation loss weighting coefficient is preset by the user. The model employs a centrally distributed distillation loss; backpropagation is performed using a joint loss function to iteratively update all trainable parameters of the backbone network and classifier until the network parameters converge, thus completing model training.

Images