Cross-modal pedestrian re-identification method based on pseudo text generation

Abstract

The invention relates to a cross-modal pedestrian re-identification method based on pseudo text generation, and belongs to the field of cross-modal pedestrian re-identification. In the feature extraction stage, global and local semantic representations of images and texts are obtained by adopting a CLIP language visual pre-training model, and weighted fusion features are generated through a feature fusion module to enhance the fine-grained representation capability. In the noise identification stage, a double-component Gaussian mixture model is provided to carry out modeling on loss distribution of image-text features, and abnormal samples in image-text matching are automatically identified and screened out. And then, reasoning the noise image by using a multi-modal large language model, and generating a pseudo text on line to replace original noise description so as to realize data enhancement and text repair. In the optimization stage, a novel triple cross-modal alignment loss function is introduced, and explicit modeling and alignment are carried out on the semantic relation between the text and the image. The whole method realizes the unification of data cleaning, semantic completion and robust training.

Description

Technical Field

[0001] The present invention belongs to the field of cross-modal person re-identification and relates to a cross-modal person re-identification method based on pseudo-text generation. Background Art

[0002] Person Re-identification (ReID) is a core technology in the field of computer vision, aiming to achieve target retrieval through cross-camera and cross-time-space pedestrian feature matching, and has important application values in fields such as public security and smart cities. With the popularization of video surveillance networks, traditional manual retrieval methods are difficult to cope with the challenges of massive data, and deep learning-based person re-identification technology has become the mainstream solution. However, problems such as perspective changes, lighting differences, and occlusion interference in actual scenarios still restrict the robustness of the system. In recent years, Text-based Person ReID has attracted much attention because it can use natural language descriptions for target retrieval, especially suitable for scenarios where witnesses can only provide text clues (such as searching for missing persons, criminal investigation tracking). This technology is essentially cross-modal image-text retrieval, and it is necessary to map text descriptions and image features to the same semantic space to achieve accurate matching.

[0003] Existing methods mainly rely on vision-language pre-training models (such as CLIP) to extract image-text features and perform cross-modal alignment through attention mechanisms or joint embedding spaces. However, these methods implicitly assume that the image-text pairs in the training data are perfectly aligned, ignoring the noise problems that are prevalent in real-world scenarios. For example:

[0004] Manual annotation is vulnerable to subjectivity and annotation errors, resulting in partial mismatch between text descriptions and image content (such as incorrect attribute descriptions, detail omissions);

[0005] The text may lack fine-grained features (such as clothing textures, accessory details), making it difficult to distinguish similar pedestrians;

[0006] The local feature interaction between images and texts is insufficient, and global features are difficult to capture subtle differences.

[0007] Specifically, the existing technologies have the following defects:

[0008] Noisy image-text pairs will cause the model to learn incorrect cross-modal associations and reduce the generalization ability;

[0009] Traditional methods rely on global feature alignment and ignore the complementarity of local details;

[0010] There is a lack of correction strategies for noisy data and it is impossible to dynamically optimize low-quality descriptions.

[0011] Although some studies have attempted to mitigate the noise problem through design or data cleaning, there is still a lack of systematic solutions. Therefore, there is an urgent need for a person re-identification method that can identify noisy data, generate fine-grained text descriptions, and improve the cross-modal alignment accuracy through multi-granularity feature fusion to address the challenges of complex real-world scenarios. Summary of the Invention

[0012] In view of this, the purpose of the present invention is to provide a cross-modal person re-identification method based on pseudo-text generation. First, the CLIP model is used as the backbone network for image and text feature extraction. Then, a noise recognition template is designed to identify noisy image-text pairs in the dataset. Next, a multi-modal large language model is used to generate fine-grained text descriptions to replace the low-quality or incorrect descriptions in the noisy image-text pairs. At the same time, a local feature fusion module is introduced to obtain finer-grained feature interactions and achieve cross-modal alignment. Finally, a full-process solution from noise detection, filtering to enhancement and compensation is constructed, aiming to improve the retrieval accuracy and generalization ability of the model in complex real-world scenarios.

[0013] To achieve the above purpose, the present invention provides the following technical solutions:

[0014] A cross-modal person re-identification method based on pseudo-text generation, comprising the following steps:

[0015] S1: Extract image features and text semantic features through the CLIP model;

[0016] S2: Perform fine-grained interaction on the extracted features through a local feature fusion module to generate enhanced global features;

[0017] S3: Use the Gaussian mixture model GMM to model the image-text pair loss value to identify noisy data;

[0018] S4: Invoke a multi-modal large language model to generate fine-grained corrected text;

[0019] S5: Use the triplet loss TAL to optimize the model and complete the full-process processing of "noise recognition - generation of correction - robust training".

[0020] Further, the S1 includes:

[0021] S11: Image feature extraction

[0022] Given the input image I ∈ R H×W×C Divided into N non-overlapping image patches of 16×16 pixels {I i | i = 1, 2, …, N}, the image patches are mapped into embedding vectors through a linear projection layer F, and position embeddings are added The finally extracted image feature representation is: f i v = [Icls ; F(I1); F(I2); …; F(I p ) + P, after being encoded by L layers of Transformer, extract the global features corresponding to the [CLS] token

[0023] S12: Text feature extraction

[0024] For the input text T, perform word segmentation and encoding using lowercase byte pair encoding with a vocabulary of 49152. The segmented text description inserts [SOS] and [EOS] embedding vectors at the beginning and end respectively to identify the start and end of the text description sentence, and then adds position encoding in the same way That is: f i t = [E(T sos ); E(T1); E(T2); …; E(T eos )] + P; Finally, project the output text feature embedding vectors into the image - text joint embedding space in the last layer of the transformer, using the vector represented by [EOS] As the global text feature representation.

[0025] Furthermore, the S2 includes:

[0026] S21: Local feature screening

[0027] Select K high - information features from the self - attention weights of the last layer of Transformer for images and texts:

[0028]

[0029] S22: Feature transformation and fusion

[0030] Enhance the screened features through MLP and the fully connected layer FC, and perform L2 normalization:

[0031]

[0032] S23: Enhanced global feature generation

[0033] Concatenate the fused features and with the original local features to generate the final feature representation.

[0034] Furthermore, the S2 also includes a resolution adaptation strategy: when the input image resolution does not match the pre - trained model, use linear 2D interpolation to adjust the position embedding, that is: P adjusted = Interp2D(P original , H new , Wnew ), where H new , W new is the target resolution.

[0035] Furthermore, in the above S3, the noise recognition includes:

[0036] S31: Calculate the contrast loss value L of the text-image pair and normalize it:

[0037]

[0038] S32: Fit a two-component Gaussian mixture model GMM:

[0039]

[0040] S33: Divide the noise data based on the posterior probability threshold:

[0041] Furthermore, in the above S4, the text correction includes:

[0042] S41: Pre-define a multi-attribute template library For example:

[0043] t k = 'Wearing [clothing], [gender] has [hair] and carries [belongings]"

[0044] Randomly select one of the templates {template} and insert it into the static instruction to obtain the following dynamic instruction:

[0045] P = "Generate a description about the overall appearance of the person, including clothing, shoes, hairstyle, gender, and belongings, in a style similar to the template: '{template}'. If some requirements in the template are not visible, you can ignore them. Do not imagine any contents that are not in the image."

[0046] S42: Input the dynamic instruction and the image of the noise text-image pair into the multi-modal large model (Qwen-VL) simultaneously to generate the corrected text, that is:

[0047] NewCaption = Ψ model (Image, P).

[0048] Furthermore, in S5, the triplet loss TAL is defined as:

[0049]

[0050] where m represents the margin value of the difference between the positive sample and the negative sample, and f v i and f t i respectively represent the image and text of a batch of samples, and f t j , f t i are the hardest negative sample, negative sample, and positive sample in a batch of samples respectively.

[0051] . A system of an implementation method includes: a feature extraction module: configuring a CLIP model, an image-text chunk processor, and a position embedding adjuster; a fusion module: including a local feature filter, an MLP / FC transformer; a noise processing module: a GMM classifier, a template generator; a training module: a TAL loss calculator, a dynamic weighting controller.

[0052] An electronic device includes a processor and a memory, and the memory stores a computer program, which implements the cross-modal person re-identification method based on pseudo-text generation when the program is executed by the processor.

[0053] A computer-readable storage medium stores computer instructions thereon, and the computer instructions implement the cross-modal person re-identification method based on pseudo-text generation when executed by a processor

[0054] The beneficial effects of the present invention are as follows: The present invention proposes a cross-modal pedestrian re-identification method based on pseudo-text generation. By introducing techniques such as noise recognition, pseudo-text generation, and triplet alignment loss, the robustness and accuracy of the model in complex scenarios are significantly improved. To more precisely explore the corresponding relationships between modalities, we introduce local features to perform finer-grained interactions between the two modalities, and use the combination of global features and local features as the final pedestrian features to enhance the expressive power of the fused features. In the early training of the neural network, clean data often has a smaller loss value than noisy data, and the deep network tends to learn clean samples faster than noisy samples. We use a Gaussian mixture model to fit the loss results, enabling the model to effectively identify noisy data. To perform data augmentation on the identified noise, we use MLLM to generate captions for the images of the noisy image-text pairs, generating more accurate and rich text descriptions. Compared with existing methods, the present invention can effectively process noisy data, improve the accuracy of cross-modal image-text matching, and has broad application prospects.

[0055] The present invention has conducted a large number of experiments on real-world image and text datasets. By comparing standard information retrieval evaluation metrics such as the R@K value, in the text-to-image retrieval task, the model of the present invention achieved 75% for R@1, 90% for R@5, and 94% for R@10 on the CUHK-PEDES dataset. On the ICFG-PEDES dataset, R@1 reached 68%, R@5 reached 83%, and R@10 reached 89%. These results reflect the accuracy of the present invention in cross-modal retrieval. Compared with existing methods, such as the IRRA and CFine methods, the present invention improved by approximately 4% and 8% respectively in R@1 for text-to-image retrieval, by approximately 2% and 5% respectively in R@5, and by approximately 2% and 4% respectively in R@10. The advanced retrieval performance of the present invention can be used for real-time prediction and meet the large-scale pedestrian re-identification retrieval requirements.

[0056] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. Brief Description of the Drawings

[0057] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0058] Figure 1 This is the schematic diagram of the present invention. Detailed Embodiments

[0059] The embodiments of the present invention will be described below through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0060] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0061] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0062] As Figure 1 shown in the schematic diagram of the overall framework structure of a cross-modal person re-identification method based on pseudo-text generation, which includes the following steps:

[0063] S1. Obtain text and image features based on the CLIP pre-trained model, and use cosine similarity to calculate the global feature similarity matrix.

[0064] S2. Select local features with higher information content from the weighting coefficients for fusion. These weighting coefficients can be directly extracted from the self-attention weights of the last Transformer module. Select the top K local text and image features according to the weighting coefficients for feature transformation, and finally obtain a more expressive fused feature embedding.

[0065] S3. In the early training of the neural network, clean data often has a smaller loss value than noisy data. The deep network tends to learn clean samples faster than noisy samples, and clean data often has a smaller loss value than noisy data. Therefore, we use two two-component Gaussian mixture models (GMMs) to fit the triplet alignment loss of each batch of samples calculated from the global features and the fusion features, and model the image-text similarity as a clean distribution and a noisy distribution. Specifically, the Gaussian component with a lower mean is assigned to the clean set, and the other is assigned to the noisy set.

[0066] S4. After the model training tends to be stable, the images in the noisy image-text pairs classified in the previous epoch are input into the MLLM, and image caption generation is performed through dynamic instructions. The generated text will replace the noisy text in the original dataset and continue the subsequent model training.

[0067] S5. Apply the global features and the fusion features to the backward feedback process during training. Use the triplet loss to iteratively optimize the model, calculate the similarity through the loss function, and sort to obtain the re-identification result.

[0068] The person re-identification task based on text description requires the model to learn more accurate and discriminative pedestrian representations from the text description. The core tasks are: text description feature extraction and content correlation measurement between different modality data. Usually, the model uses the text as a query to retrieve the relevant image data.

[0069] Due to factors such as environmental interference, inaccurate language descriptions, mislabeling, or missing information, noise is common, which may lead to some inaccurate / mismatched text descriptions of images. If there are too many such imperfect / mismatched image-text pairs, it will cause overfitting problems and reduce the performance of existing person re-identification methods. Therefore, the present invention designs a person re-identification strategy based on noise processing. At the same time, by simultaneously using the global features and local features of feature extraction, the correlation between images and texts is enhanced, thereby improving the retrieval accuracy in the cross-modal person re-identification task.

[0070] In S1 of this embodiment, this example uses the image encoder Vit of clip to extract visual features. In terms of text, it uses the text encoder CLIP-Xformer of clip to extract text features. Given the input image I ∈ R H×W×C , the image is evenly divided into N non-overlapping image patches of 16×16 pixels. Then, the N image patches are projected through a linear projection layer to obtain N D-dimensional image patch vectors, and a learnable [CLS] embedding vector and position embedding are inserted before the input image patch vector sequence, that is:

[0071] f i v= [I cls ; F(I1); F(I2); …; F(I p )] + P

[0072] where, f i v represents the embedding vector of the image input sequence, and F represents the linear projection layer that maps the patch image to a D-dimensional vector. Then the image feature embedding vector input into the ViT network can be expressed as: After passing through L layers of L-layer transformer blocks, the embedding vector will be mapped to the image-text joint embedding space using linear projection, where is used as the global image feature representation.

[0073] For the input text T, [SOS] and [EOS] embedding vectors are inserted at the beginning and end respectively to identify the beginning and end of the text description sentence. And to ensure computational efficiency, the maximum length of the text description sequence is set to 77. To learn the relative position relationship of words in the sentence, the position embedding will also be added to the word vector input sequence. After the projection transformation, the input sequence can be expressed as:

[0074] f i t = [E(T sos ) ; E(T1); E(T2); …; E(T eos )] + P

[0075] where D represents the embedding vector of the text input sequence, and E represents the projection layer that maps the word vector to a D-dimensional vector. The final text feature embedding vector input into the transformer can be expressed as:

[0076]

[0077] Similarly, finally, in the last layer of the transformer, the output text feature embedding vector is projected into the image-text joint embedding space, and the vector represented by [EOS] is used as the global text feature representation. To calculate the similarity between any image-text pair (I, T), we directly use the global features [CLS] and [SOS] of the image and text, and obtain the global feature embedding (gfe) similarity through cosine similarity, that is:

[0078]

[0079] In S2 of this embodiment, the purpose of the feature fusion module is to more precisely explore the correspondence between modalities for finer-grained interaction between the two modalities.

[0080] Since the global feature is obtained by weighted aggregation of all local features and the weighting coefficients reflect the correlation between the global feature and each local feature, we can use tokens with larger amounts of information to learn more discriminative embedding representations to obtain more representative global features. We can select local features with higher information content from the weighting coefficients for fusion, select the top K local features of the image and text according to the weighting coefficients for feature transformation, and finally obtain a more expressive fused feature embedding (token fusion embedding, tfe). The feature transformation uses the same embedding module as the residual block, that is:

[0081]

[0082] Among them, represents the top K local features with the largest amount of information selected, are the image and text features after L2-norm normalization. Finally, we use and to calculate the cosine similarity S of feature fusion tfe , and together with the global feature similarity S gfe as the similarity of the image-text pair for cross-modal alignment training and inference.

[0083] In S3 of this embodiment, in the early training of the neural network, clean data often has a smaller loss value than noisy data. The deep network tends to learn clean samples faster than noisy samples, and clean data often has a smaller loss value than noisy data. Therefore, we use two two-component Gaussian mixture models GMM to fit the triplet alignment loss of each batch of samples calculated from gfe and tfe, and model the image-text similarity as a clean distribution and a noise distribution, that is, assign the Gaussian component with a lower average value to the clean set and the other to the noise set respectively.

[0084] The triplet TAL loss for the i-th batch of samples is l i , input this loss into the GMM, optimize it using the expectation-maximization algorithm, and calculate the posterior probability:

[0085] p(k|l i ) = p(k)p(l i |k) / p(l i )

[0086] Among them, the posterior probability p(k|l i), where \(k\in\{0,1\}\) represents the probability that the \(i\)-th sample is classified as a clean or noisy sample pair. By setting a threshold \(\delta = 0.5\), we can divide the dataset containing \(M\) text-image pairs into a clean dataset and a noisy dataset

[0087] In step S4 of this embodiment, the open-source Qwen-VL has obvious advantages over the current best large generation models in multiple vision-language tasks, and its capabilities cover a more comprehensive range. Moreover, its deployment in the environment is very lightweight and convenient. Therefore, we choose the Qwen-VL model for pseudo-text generation.

[0088] Inputting \(P\) to the multi-modal large model MLLM can infer and generate a text description of an accurate pedestrian image. However, when generating text descriptions with simple static instructions, similar sentence pattern problems are likely to occur, leading to model overfitting and restricting its generalization ability. Therefore, a diverse text generation method based on template dynamic instructions is used. We randomly select one of the templates and insert it into the static instruction, and then input the resulting dynamic instruction \(P\) and the image in the noisy text-image pair into the MLLM to obtain the text description, that is:

[0089] NewCaption = Ψ model (Image, P)

[0090] In subsequent training, use the newly generated text to replace the noisy text in the noisy text-image pair for subsequent training.

[0091] In step S5 of this embodiment, the commonly used loss functions for person re-identification include InfoNCE, triplet ranking loss (TRL), and image-text similarity distribution matching (SDM) loss. However, these loss functions have problems such as the difficulty in choosing the hardest negative sample and the similarity margin. Especially when dealing with person text re-identification in a noisy environment, it may lead to local optimality or even model collapse due to over-focusing on the hardest negative sample. But the triplet alignment loss (TAL) alleviates the risk of over-optimization of the hardest negative sample by imposing an upper bound constraint on all negative samples, making the training more stable and comprehensive. When introducing the temperature coefficient \(\tau\), TAL comprehensively considers all positive and negative sample pairs, relaxes the optimization of the hardest negative sample, and thus makes the training more stable and comprehensive. That is:

[0092]

[0093] where \(m\) represents the difference margin value between the positive sample and the negative sample, and \(f\) t i respectively represent the image and text of a batch of samples, and \(f\) t j, f t i They are the most difficult negative sample, negative sample, and positive sample in a batch of samples respectively.

[0094] The network model undergoes multiple rounds of training. The training process should be to extract features and then continuously optimize the model parameters by feeding the triplet loss back to the model. Therefore, the recognition process actually calculates the Euclidean distance between the input image and each image in the image library after the model is trained through the loss function, sorts the images in the library in ascending order of distance, and thus obtains the top n images with the smallest distance in the image library, that is, the images with the highest similarity, denoted as Rank-k, which refers to the ratio of the number of correctly retrieved images to the number of all images that should be retrieved. Specifically, it represents the probability of correctly identifying a pedestrian in the top n query results returned on a preset test set, where the common ones are 1, 5, and 10.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A cross-modal pedestrian re-identification method based on pseudo-text generation, characterized in that: It includes the following steps: S1: Extract image features and text semantic features through the CLIP model; S2: Perform fine-grained interaction on the image features and text semantic features through a local feature fusion module to generate enhanced global features; S3: Use the Gaussian mixture model GMM to model the image-text pair loss value and identify noisy image-text pairs; S4: Invoke a multi-modal large language model to generate fine-grained corrected text and replace the low-quality text in the noisy image-text pairs; S5: Optimize the enhanced global features using the triplet loss TAL to complete cross-modal person re-identification.

2. The cross-modal pedestrian re-identification method based on pseudo-text generation according to claim 1, characterized in that: The S1 includes: S11: Image feature extraction Given an input image \(I\in\mathbb{R}\) H×W×C Divide it into \(N\) non - overlapping image patches of \(16\times16\) pixels \(\{I\) i \(|i = 1,2,\cdots,N\}\), map the image patches to embedding vectors through a linear projection layer \(F\), and add position embeddings The finally extracted image features are represented as: \(f\) i v \(=[I\) cls ; \(F(I_1)\); \(F(I_2)\); \(\cdots\); \(F(I\) p )]+P\), after being encoded by \(L\) layers of Transformer, extract the global features corresponding to the \([CLS]\) token S12: Text feature extraction For the input text T, word segmentation and encoding are performed using the lowercase byte pair encoding with a vocabulary of 49152. The segmented text description inserts [SOS] and [EOS] embedding vectors at the beginning and end respectively to identify the beginning and end of the text description sentence, and then position encoding is added in the same way. That is: f i t = [E(T sos )); E(T1); E(T2); …; E(T eos ))] + P; Finally, in the last layer of the transformer, the output text feature embedding vectors are projected into the image-text joint embedding space, and the vector represented by [EOS] is used as the global text feature representation.

3. The cross-modal pedestrian re-identification method based on pseudo-text generation according to claim 2, wherein: The S2 includes: S21: Local feature screening Select K high-information features from the last-layer Transformer self-attention weights of the image and text: K = 0.3 × N, S22: Feature transformation and fusion Enhance the screened features through MLP and fully connected layer FC, and perform L2 normalization: S23: Enhanced global feature generation The fused features and are concatenated with the original local features to generate the final feature representation.

4. The cross-modal pedestrian re-identification method based on pseudo-text generation according to claim 2, characterized in that: The S2 also includes a resolution adaptation strategy: when the input image resolution does not match the pre-trained model, linear 2D interpolation is used to adjust the position embedding, that is: P adjusted = Interp2D(P original , H new , W new ), where H new , W new is the target resolution.

5. The cross-modal pedestrian re-identification method based on pseudo-text generation according to claim 3, characterized in that: In the S3, the noise identification includes: S31: Calculate the contrast loss value L of the text-image pair and normalize it: S32: Fit a two-component Gaussian mixture model GMM: S33: Divide the noise data based on the posterior probability threshold:

6. The cross-modal pedestrian re-identification method based on pseudo-text generation according to claim 1, wherein: In the S4, the text correction includes: S41: Pre-define a multi-attribute text template library to generate prompts containing dynamic instructions; S42: Input the dynamic instructions and the images of the noisy image-text pairs into the multi-modal large language model to generate corrected text.

7. The cross-modal pedestrian re-identification method based on pseudo-text generation according to claim 1, characterized in that: In the S5, the triplet loss TAL is defined as: where m represents the margin value of the difference between the positive sample and the negative sample, f v i and f t i respectively represent the images and texts of a batch of samples, and f t j , f t i are the hardest negative sample, the negative sample, and the positive sample in a batch of samples respectively.

8. A system for implementing the method according to any one of claims 1 to 6, characterized in that: It includes: A feature extraction module for extracting global features of images and texts through the CLIP model; A fusion module for screening, transforming, and splicing local features to generate enhanced features; A noise processing module including a GMM classifier and a multi-modal large language model generator; A training module for optimizing model parameters based on the triplet loss.

9. An electronic device, comprising a processor and a memory, characterized in that: The memory stores a computer program, which when executed by the processor, implements the cross-modal person re-identification method based on pseudo-text generation according to any one of claims 1 to 6.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the cross-modal person re-identification method based on pseudo-text generation according to any one of claims 1 to 6 is implemented.

Images