A dependency attribute enhanced text search image pedestrian re-identification method

Abstract

The application discloses a kind of text search image pedestrian re-identification methods of attribute dependence enhancement, comprising: text modal pre-processing is carried out to enhance the semantic features of text;Text feature T is obtained by text encoder f And introduce self-attention network Attn, to output the text feature E after dependence enhancement;Visual modal pre-processing is carried out, and pedestrian image V p Data enhancement is carried out, and then image feature F is obtained by visual encoder;Text feature E and image feature F are embedded into features, to obtain text embedding t and image embedding v;Classification loss and ranking loss are calculated to train attribute dependence enhancement text search image pedestrian re-identification model;As described above, the prior knowledge of the present application uses dependence relationship to mine attribute information to enhance semantic features, which overcomes the deficiencies of conventional image-text matching methods in the field of text search image pedestrian re-identification.

Description

Technical Field

[0001] This invention relates to a method for pedestrian re-identification in text search images, belonging to the fields of deep learning and multimodal learning. Background Technology

[0002] To improve efficiency and address extreme situations where faces cannot be effectively recognized due to dense crowds, low resolution of surveillance cameras, or unfavorable shooting angles, researchers have developed image-based person re-identification (IIPR) technology, which extracts overall pedestrian features, as an important supplement to face recognition technology. IIPR refers to automatically searching for matching pedestrian images across multiple cameras in a multi-camera network with non-overlapping view domains using the target person's image.

[0003] However, in the early stages of video surveillance, images of the target person are often lacking, making "image-to-image search" impossible. Therefore, text information provided by witnesses becomes crucial evidence for video surveillance and key clues for tracking the target. This objective is known as Text-to-Image Person Re-identification (TIPR). On one hand, TIPR helps to promptly screen target individuals and establish rapid-response security and control mechanisms. On the other hand, in crowded public areas such as airports and train stations, TIPR technology can be used to quickly locate lost children and the elderly, playing an immeasurable role in safeguarding the vital interests of the people and creating a safe environment.

[0004] TIPR aims to retrieve target people from a pedestrian image database given text. Its main challenge is learning features robust to images from free-viewpoint perspectives (such as pose, lighting, and camera viewpoint) and to free-form text, making TIPR more challenging than IIPR. One current mainstream approach utilizes attributes as a means of representing the semantics of data. Attributes, as soft biometric features in TIPR, are human-understandable semantic concepts, such as gender, age, and clothing descriptions. Due to their modality-invariant properties, attributes can serve as a good mediator between textual and visual modalities of data semantics.

[0005] Although existing attribute-based methods have achieved good performance, they are essentially variations of general cross-modal retrieval (CMR) methods and are sometimes not optimal solutions for semantically rich TIPR tasks. Existing text-to-image matching methods in the field of text-search image person re-identification have shortcomings in person attribute mining.

[0006] A method for unsupervised adaptive pedestrian re-identification, disclosed in CN202110399589.7, includes: a data acquisition module for acquiring multiple source domain sample subsets and multiple target domain sample subsets; a network model training module for obtaining the classification loss function and sample invariance loss function of the pedestrian re-identification network model, sorting and stratifying the pedestrian images in the target domain sample subsets according to the similarity between each pedestrian image in the source domain sample subsets and each pedestrian image in the target domain sample subsets to obtain a stratified loss function, and then iteratively optimizing the pedestrian re-identification network model; and a re-identification module for performing pedestrian re-identification using the optimized pedestrian re-identification network model to obtain images that are the same as or similar to the pedestrian images to be identified. This system addresses the shortcomings of conventional image-text matching methods in pedestrian attribute mining. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a text search image person re-identification method with dependency attribute enhancement. This method uses prior knowledge of dependency relationships to mine attribute information to enhance semantic features, thereby overcoming the shortcomings of conventional text-image matching methods in the field of text search image person re-identification in terms of pedestrian attribute mining.

[0008] The present invention adopts the following specific technical solution: A method for person re-identification in text search images with enhanced dependency attributes, the steps of which are as follows: S1. Perform text modality preprocessing, including using the natural language processing tool SpaCy to preprocess the raw text. Dependency analysis is performed to obtain dependency relation output. For a node of a word By calculating its parent node and Each word corresponds to a node exist Dependency distance in We obtain a side length of matrix And map it onto the Gaussian function to obtain the dependency matrix. This is used to enhance the semantic features of text; S2, will Feed to the text encoder To obtain text features Introducing self-attention networks Will and Fusion is used to achieve dependency enhancement and output the dependency-enhanced text features. Text features Depend on A person's attention Output Text features obtained by splicing The expression is:

[0009] Among them, text features Depend on A person's attention Output obtained by piecing together , , Indicates the corresponding number Each word Word embedding; S3. Perform visual modality preprocessing on pedestrian images. Perform data augmentation, transform it into a tensor, normalize it, and then standardize it to obtain information about... Diverse samples to increase Diversity; S4, will Feed to the vision encoder To obtain image features A pre-trained ResNet50 network is used to extract image features. The pre-trained ResNet50 network extracts pedestrian image features in five stages (Stage 0, Stage 1, ... Stage 4) in a progressively deeper manner. Simultaneously, the last two layers of ResNet50 (including the average pooling layer and the fully connected layer) are removed, thus preserving the image features. , , Representing image features Width and length, express The number of channels, and The number of channels is the same; S5. Text Features and image features Feature embedding is performed to obtain text embeddings. and image embedding This is used to reduce the differences between different modes; S6. Embed the text from step S5 and image embedding Feed to a fully connected layer A prediction head is created, and the probability of pedestrians is predicted using the Softmax function. The corresponding predicted pedestrian probability is , The corresponding predicted pedestrian probability is , and The expressions are as follows:

[0010]

[0011] in, This represents the total number of unique pedestrian labels in the training set. The probability of pedestrians is Each of these is a set of probability values ​​predicted for each pedestrian; S7. Calculate the classification loss (ID Loss) and ranking loss (Ranking Loss), and use the sum of the two losses to backpropagate the entire network to jointly train the attribute dependency-enhanced text search image person re-identification (ADA).

[0012] Furthermore, the specific steps of step S1 include: S1.1: Using the natural language processing tool SpaCy to... The word segmentation was performed, and the segmentation result was as follows:

[0013] in, Indicates the first One word, Indicates the number of words; S1.2: Using the natural language processing tool SpaCy to... Dependency analysis yielded ,for Each word in according to Calculate its parent word to each word Dependency distance There are arrays For the root node (ROOT), i.e., words without a parent node, The dependency distance to itself is calculated to be 0. Further calculations yield the distance from... Transformation ; S1.3: Will Mapped to Centered on, variance is Gaussian function superior:

[0014] in ,and , Indicates the first The word for the first The probability density of a pedestrian attribute based on the parent word of a given word.

[0015] Furthermore, the specific steps of step S2 include: S2.1: Will Feed to the text encoder To obtain text features Text features The expression is: ,

[0016] For the features of the i-th word, The calculation formula is:

[0017] in, , They represent the first Word vectors The forward and backward hidden states, Indicates the number of channels; S2.2: Introducing a self-attention network Will and Fusion, its self-attention network Settings A person's attention , No. A person's attention The calculation process is as follows: S2.2.1: By Obtain the relevant query vector key vector and value vector Its expression is as follows:

[0018] in, , This indicates the number of channels in the attention head. For a learnable weight matrix, further for get ; S2.2.2: To avoid Chinese index Conflict of use, Corresponding to Calculate the corresponding dot product and divide by scaling factor , to obtain a representation of the first Each word For the first Each word Attention score The calculation formula is as follows:

[0019] S2.2.3: Then, the attention weight matrix is ​​derived. ,and ,Will and Perform the Hadamard product and map it to the range of 0 to 1 using the Softmax function. and have:

[0020] in yes right The new attention score obtained after the intervention, therefore, ; S2.2.4: In Attention output is obtained in the process ,Will A person's attention Output Text features are obtained by splicing .

[0021] Furthermore, in step S2.2.4, attention output The calculation formula is: .

[0022] Furthermore, the data enhancement technique for the pedestrian image in step S3 is at least one of the following: horizontal flipping, edge filling, grayscale transformation, statistical cropping, random cropping, random cropping and re-stitching, random slight angle rotation, random position transformation, random noise addition, random erasure, and image sharpening.

[0023] Furthermore, the specific steps of step S5 are as follows: S5.1: In In this dimension, and Perform max pooling to obtain the original text embeddings. and image embedding ; S5.2: will They all went through a Shared weight matrix Matrix multiplication is performed to bridge the semantic gap between the text and visual modalities, resulting in the following formulas for calculating the final text embedding t and image embedding v:

[0024] in , This indicates the number of new channels.

[0025] Furthermore, the specific steps of step S7 are as follows: S7.1 calculates the classification loss (ID Loss), first given the pedestrian label. And perform one-hot encoding to obtain: ; The formula for calculating its classification loss, ID Loss, is:

[0026] S7.2: Ranking Loss is calculated based on Triple Loss using the following formula:

[0027] in, For positive sample pairs, For randomly selected negative samples, the pedestrian labels are different. , Let be the cosine similarity.

[0028] Furthermore, the cosine similarity The calculation formula is: .

[0029] Furthermore, the method is verified and analyzed through experiments.

[0030] The beneficial effects of this invention are as follows: The beneficial effects of the dependency attribute-enhanced text search image person re-identification method designed in this invention are as follows: (1) In the text feature extraction stage, this invention is adapted to pedestrian re-identification in text search images. In order to fully explore the rich attribute features in pedestrian text, attribute dependency enhancement is further explicitly implemented for text features. Specifically, attributes are mined by focusing on words with deeper dependency relationships; (2) For and For feature processing of different modalities, a lightweight and efficient feature embedding approach is adopted to improve the speed of modal feature processing. In addition, the transformation process from [previous method name] to [previous method name] is parameterless, thus not affecting the overall processing speed of the method. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0032] Figure 1 This is an overview diagram of attribute-dependent enhanced text search image pedestrian re-identification methods; Figure 2 This is a schematic diagram of the process of fusing dependency matrices based on self-attention networks; Figure 3 This is a scatter plot of the parts of speech of the text in the CUHK-PEDES and ICFG-PEDES datasets; Figure 4 It is a structural diagram of the attribute dependency tree of a piece of text. Detailed Implementation

[0033] The following examples further explain and clarify the invention, but the specific examples do not limit the invention in any way.

[0034] Example 1 A dependency attribute-enhanced method for pedestrian re-identification in text search images, such as... Figures 1-2 As shown, the steps of the method are as follows: S1. Perform text modality preprocessing, including using the natural language processing tool SpaCy to preprocess the raw text. Dependency analysis is performed to obtain dependency relation output. For a node of a word By calculating its parent node and Each word corresponds to a node exist Dependency distance in We obtain a side length of matrix And map it onto the Gaussian function to obtain the dependency matrix. This is used to enhance the semantic features of text; S2, will Feed to the text encoder To obtain text features Introducing self-attention networks Will and Fusion is used to achieve dependency enhancement and output the dependency-enhanced text features. Text features Depend on A person's attention Output Text features obtained by splicing The expression is:

[0035] Among them, text features Depend on A person's attention Output obtained by piecing together , , Indicates the corresponding number Each word Word embedding; S3. Perform visual modality preprocessing on pedestrian images. Perform data augmentation, transform it into a tensor, normalize it, and then standardize it to obtain information about... Diverse samples to increase Diversity; S4, will Feed to the vision encoder To obtain image features A pre-trained ResNet50 network is used to extract image features. The pre-trained ResNet50 network extracts pedestrian image features in five stages (Stage 0, Stage 1, ... Stage 4) in a progressively deeper manner. Simultaneously, the last two layers of ResNet50 (including the average pooling layer and the fully connected layer) are removed, thus preserving the image features. , , Representing image features Width and length, express The number of channels, and The number of channels is the same; S5. Text Features and image features Feature embedding is performed to obtain text embeddings. and image embedding This is used to reduce the differences between different modes; S6. Embed the text from step S5 and image embedding Feed to a fully connected layer A prediction head is created, and the probability of pedestrians is predicted using the Softmax function. The corresponding predicted pedestrian probability is , The corresponding predicted pedestrian probability is , and The expressions are as follows:

[0036]

[0037] in, This represents the total number of unique pedestrian labels in the training set. The probability of pedestrians is Each of these is a set of probability values ​​predicted for each pedestrian; S7. During the training phase, the classification loss (ID Loss) between the predicted and true labels is calculated, as is the ranking loss. The sum of these two losses is then used for backpropagation of the entire network. The backpropagation of Attn is sensitive to pedestrian attributes, ultimately achieving joint training of Attribute Dependency Enhanced Text Search Image Person Re-identification (ADA). When calculating the classification loss ID, given the pedestrian label... And perform one-hot encoding to obtain ; The formula for calculating its classification loss, ID Loss, is:

[0038] The ranking loss, calculated based on the triple loss formula, is as follows:

[0039] in, For positive sample pairs, For randomly selected negative samples, the pedestrian labels are different. , It is a cosine similarity calculation; cosine similarity The calculation formula is: .

[0040] Furthermore, the specific steps of step S1 include: S1.1: Using the natural language processing tool SpaCy to... The word segmentation was performed, and the segmentation result was as follows:

[0041] in, Indicates the first One word, Indicates the number of words; S1.2: Using the natural language processing tool SpaCy to... Dependency analysis yielded ,for Each word in according to Calculate its parent word to each word Dependency distance There are arrays For the root node (ROOT), i.e., words without a parent node, The dependency distance to itself is calculated to be 0. Further calculations yield the distance from... Transformation ; S1.3: Will Mapped to Centered on, variance is Gaussian function superior:

[0042] in ,and , Indicates the first The word for the first The probability density of a pedestrian attribute based on the parent word of a given word.

[0043] Furthermore, the specific steps of step S2 include: S2.1: Will Feed to the text encoder To obtain text features Text features The expression is: ,

[0044] For the features of the i-th word, The calculation formula is:

[0045] in, , They represent the first Word vectors The forward and backward hidden states, Indicates the number of channels; S2.2: Introducing a self-attention network Will and Fusion, its self-attention network Settings A person's attention , No. A person's attention The calculation process is as follows: S2.2.1: By Obtain the relevant query vector key vector and value vector Its expression is as follows:

[0046] in, , This indicates the number of channels in the attention head. For a learnable weight matrix, further for get ; S2.2.2: To avoid Chinese index Conflict of use, Corresponding to Calculate the corresponding dot product and divide by scaling factor , to obtain a representation of the first Each word For the first Each word Attention score The calculation formula is as follows:

[0047] S2.2.3: Then, the attention weight matrix is ​​derived. ,and ,Will and Perform the Hadamard product and map it to the range of 0 to 1 using the Softmax function. and have:

[0048] in yes right The new attention score obtained after the intervention, therefore, ; S2.2.4: In Attention output is obtained in the process ,Will A person's attention Output Text features are obtained by splicing .

[0049] Furthermore, in step S2.2.4, attention output The calculation formula is: .

[0050] Furthermore, the data enhancement technique for the pedestrian image in step S3 is at least one of the following: horizontal flipping, edge filling, grayscale transformation, statistical cropping, random cropping, random cropping and re-stitching, random slight angle rotation, random position transformation, random noise addition, random erasure, and image sharpening.

[0051] Furthermore, the specific steps of step S5 are as follows: S5.1: In In this dimension, and Perform max pooling to obtain the original text embeddings. and image embedding ; S5.2: will They all went through a Shared weight matrix Matrix multiplication is performed to bridge the semantic gap between the text and visual modalities, resulting in the following formulas for calculating the final text embedding t and image embedding v:

[0052] in , This indicates the number of new channels.

[0053] In a specific application of this embodiment, for a given segment of original text... Acquired from the camera Image of a pedestrian Extract the embedded text respectively and Image embedding ,calculate right cosine similarity is obtained From the given metrics, the top ten pedestrian images with the highest metrics are selected as the most suitable. The search results described enable text-to-image pedestrian re-identification.

[0054] In summary, this embodiment, in the text feature extraction stage, is suitable for pedestrian re-identification in text search images. To fully exploit the rich attribute features in pedestrian text, it further explicitly enhances attribute dependencies on text features. Specifically, it mines attributes by focusing on words with deeper dependency relationships; secondly, this embodiment... and For feature processing of different modalities, a lightweight and efficient feature embedding approach is adopted to improve the speed of modal feature processing. In addition, the transformation process from [previous method name] to [previous method name] is parameterless, thus not affecting the overall processing speed of the method.

[0055] Example 2 To verify the effectiveness and accuracy of this invention, this embodiment 2 conducts experiments on the publicly available mainstream datasets CUHK-PEDES and ICFG-PEDES. The training set of the datasets is trained several times in batches. On the test set, Rank-1, Rank-5, and Rank-10 are used as evaluation metrics. Rank-1 represents the probability that the first image in the query results and the search image belong to the same pedestrian. Similarly, Rank-5 and Rank-10 represent the probabilities that the first five and first ten images in the query results and the search image belong to the same pedestrian.

[0056] Experimental parameters: Batch size was 64, number of training epochs was 60, initial learning rate was 0.001, and the Adam method was used as the optimizer.

[0057] This embodiment 2 compares the method of the present invention described in embodiment 1 with a series of advanced TIPR methods in recent years on two mainstream datasets. Specific experimental results are shown in Table 1. The table shows that previous methods achieved good accuracy, but these LSTM or Transformer-based methods still lack attention to the deep structural details of text descriptions. The present invention uses dependency relations as a data augmentation technique in the text modality, explicitly intervening in the focus on attributes to extract key clues. The present method achieves higher accuracy on Rank-1 of CUHK-PEDES and ICFG-PEDES, and also shows good competitiveness on Rank-5 and Rank-10.

[0058]

[0059] Table 1 Example 3 To verify the effectiveness and accuracy of this invention, Example 3 uses the method described in Example 1 to conduct experiments on the publicly available mainstream datasets CUHK-PEDES and ICFG-PEDES. Experimental parameters: batch size of 64, number of training epochs of 60, initial learning rate of 0.001, and the Adam method as the optimizer.

[0060] This embodiment 3 explains the effectiveness of attribute dependency enhancement from the perspective of the relationship between part-of-speech and dependency distance. First, as... Figure 4As shown, text descriptions are extracted from the CUHK-PEDES and ICFG-PEDES datasets. Dependency analysis is performed using SpaCy. For each text description, a dependency tree is created. The dependency distance is calculated as the number of nodes traversed from the parent node to the ROOT node for each word, yielding the relative dependency distance relative to the text description length. Then, a table is maintained to record the total number of each part of speech and the sum of the relative dependency distances. Finally, the sum of the relative dependency distances is divided by the total number of corresponding parts of speech to obtain the average dependency distance, which is then plotted as a scatter plot. Figure 3 , Figure 3 The x-axis represents quantity and the y-axis represents average dependency distance. Figure 3 It can be clearly observed that while the modifier attribute (ADJ) and noun attribute (NOUN) have the largest number of words, they also have a high dependency depth. This indicates that when the intervention model using the attribute dependency enhancement strategy focuses on the deeper dependency structure of the text, it often focuses on the modifier attribute and noun attribute. This further proves that using pedestrian attributes to represent the semantics of the data in text search image pedestrian re-identification is beneficial for extracting semantic features.

[0061] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for pedestrian re-identification in text search images with enhanced dependency attributes, characterized in that, The steps of the method are as follows: S1. Perform text modality preprocessing, including using the natural language processing tool SpaCy to preprocess the raw text. Dependency analysis is performed to obtain dependency relation output. By calculation Dependency distance in , to obtain the matrix And map it onto a Gaussian function to obtain the dependency matrix. This is used to enhance the semantic features of the text; S1.1: Using the natural language processing tool SpaCy to... The word segmentation was performed, and the segmentation result was as follows: in, Indicates the first One word, Indicates the number of words; S1.2: Using the natural language processing tool SpaCy to... Dependency analysis yielded ,for Each word in according to Calculate its parent word to each word Dependency distance There are arrays For the root node (ROOT), i.e., words without a parent node, The dependency distance to itself is calculated to be 0. Further calculations yield the distance from... Transformation matrix ; S1.3: Will Mapped to Centered on, variance is Gaussian distribution The above yields the dependency matrix. for: in ,and , Indicates the first The word for the first The probability density of a pedestrian attribute based on the parent word of a given word; S2, will Feed to the text encoder To obtain text features Introducing self-attention networks Will and Fusion is used for dependency enhancement, outputting dependency-enhanced text features. Its text features The calculation formula is: Among them, text features Depend on A person's attention Output obtained by piecing together , , Indicates the corresponding number Each word Word embedding; S3. Preprocess the visual modalities for pedestrian images. Perform data augmentation, transform it into a tensor, normalize it, and then standardize it to obtain information about... Diverse samples ; S4, will Feed to the vision encoder To obtain image features ; S5. Text Features and image features Feature embedding is performed to obtain text embeddings. and image embedding This is used to reduce the differences between different modes; S6. Embed the text from step S5 and image embedding Feed to a fully connected layer A prediction head is created, and the probability of pedestrians is predicted using the Softmax function. The corresponding predicted pedestrian probability is , The corresponding predicted pedestrian probability is , and The expressions are as follows: in, This represents the total number of unique pedestrian labels in the training set. The probability of pedestrians is The pedestrian probability is a set of predicted probability values ​​for each pedestrian; S7. Calculate the classification loss (ID Loss) and ranking loss (Ranking Loss), and use the sum of the two losses to backpropagate the entire network to jointly train the attribute dependency-enhanced text search image person re-identification model.

2. The method for person re-identification in text search images with enhanced dependency attributes according to claim 1, characterized in that, The specific steps of step S2 include: S2.1: Will Feed to the text encoder To obtain text features Text features The expression is: ， in, For the first Features of each word The calculation formula is: in, , They represent the first Word vectors The forward and backward hidden states, Indicates the number of channels; S2.2: Introducing a self-attention network Will and Fusion, its self-attention network Settings A person's attention , No. A person's attention The calculation process is as follows: S2.2.1: By Obtain the relevant query vector key vector and value vector Its expression is as follows: in, , This indicates the number of channels in the attention head. For a learnable weight matrix, further for get ; S2.2.2: To avoid Chinese index Conflict of use, Corresponding to ,calculate dot product and divide by scaling factor , to obtain a representation of the first Each word For the first Each word Attention score ,That The calculation formula is as follows: S2.2.3: Then, the attention weight matrix is ​​derived. ,and ,Will and Perform the Hadamard product and map it to the range of 0 to 1 using the Softmax function. and have: in, yes right The new attention score obtained after the intervention, therefore, ; S2.2.4: In Attention output is obtained in the process ,Will A person's attention Output Text features are obtained by splicing .

3. The method for person re-identification in text search images with enhanced dependency attributes according to claim 2, characterized in that, In step S2.1 The solution process is as follows: Step 1: Use Word2Vec to... Each word in Mapped to the corresponding word vectors ,get: Step 2: Feed to BiLSTM processing from arrive as well as arrive The word vectors, for word vectors have:

4. The method for person re-identification in text search images with enhanced dependency attributes according to claim 2, characterized in that, Attention output in step S2.2.4 The calculation formula is:

5. The method for person re-identification in text search images with enhanced dependency attributes according to claim 1, characterized in that, The data enhancement technique for the pedestrian image in step S3 is at least one of the following: horizontal flipping, edge filling, grayscale transformation, statistical cropping, random cropping, random cropping and stitching, random slight angle rotation, random position transformation, random noise addition, random erasure, and image sharpening.

6. The method for person re-identification in text search images with enhanced dependency attributes according to claim 1, characterized in that, The specific steps of step S5 are as follows: S5.1: In In this dimension, and Perform max pooling to obtain the original text embeddings. and image embedding ; S5.2: will They all went through a Shared weight matrix Matrix multiplication is performed to bridge the semantic gap between the text and visual modalities, resulting in the following formulas for calculating the final text embedding t and image embedding v: in , This indicates the number of new channels.

7. The method for person re-identification in text search images with enhanced dependency attributes according to claim 1, characterized in that, The specific steps of step S7 are as follows: S7.1 calculates the classification loss (ID Loss), first given the pedestrian label. And perform one-hot encoding to obtain: ； The formula for calculating its classification loss, ID Loss, is: S7.2: Ranking Loss is calculated based on Triple Loss using the following formula: in, For positive sample pairs, For randomly selected negative samples, the pedestrian labels are different. , Let be the cosine similarity.

8. The method for person re-identification in text search images with enhanced dependency attributes according to claim 7, characterized in that, The cosine similarity The calculation formula is: in, express The transpose of .

9. The method for person re-identification in text search images with enhanced dependency attributes according to claim 1, characterized in that, The method was verified and analyzed experimentally.

Images