A method and system for cross-domain open set person re-identification based on unsupervised domain adaptation

Abstract

This invention relates to a method and system for cross-domain open set person re-identification based on unsupervised domain adaptation. The method includes: constructing a cross-domain feature adversarial alignment module including a feature extractor and a cross-domain discriminator; constructing an integrated binary classification bias module including multiple binary classification bias modules; alternately optimizing the cross-domain feature adversarial alignment module and the integrated binary classification bias module in an adversarial manner by combining multi-class cross-entropy loss, difference loss, and bias loss in the source and target domains; after the model converges, determining whether the person in the current test image belongs to a known object or an unknown object based on the re-identification result and bias score output by the model; this invention uses a joint training method to simultaneously perform cross-domain feature distribution adversarial alignment and unknown object identification, ensuring efficient re-identification of known objects while also identifying unknown objects with high accuracy.

Description

Technical Field

[0001] This invention can be applied to the field of cross-domain person re-identification, and in particular relates to an open set person re-identification method and system based on unsupervised domain adaptation. Background Technology

[0002] To meet residents' needs for public safety, massive video surveillance systems are widely used in public places, providing a wealth of video and image data for target tracking and criminal investigations. However, due to blind spots and the limited shooting range of individual cameras, manually reviewing surveillance footage to track the same target across cameras is time-consuming and labor-intensive. Person re-identification technology aims to automatically identify the same target across different cameras using deep learning, pattern recognition, and other technologies, thereby improving the timeliness and accuracy of analyzing surveillance data and facilitating smart city construction, criminal investigations, and the search for missing persons. Traditional person re-identification technologies based on manually designed person representations have poor cross-domain generalization performance and high algorithm complexity, making them unsuitable for today's scenarios with massive amounts of surveillance video and images.

[0003] In recent years, the improvement of computing power has promoted the development of deep learning technology. In the field of person re-identification, convolutional neural networks have achieved an overwhelming advantage over traditional manually designed feature methods due to their powerful feature extraction capabilities. However, in today's massive video and image scenarios, the data labeling cost of supervised person re-identification technology based on deep learning is high. In order to reduce the data labeling cost and maintain high accuracy in cross-domain person re-identification, unsupervised cross-domain person re-identification technology has become a research hotspot. Since there are domain differences such as shooting angle, brightness, person pose, and background between videos or images captured by different cameras, unsupervised cross-domain person re-identification technology achieves the generalization of model recognition ability by reducing the domain differences between the richly labeled source domain dataset and the unlabeled target domain dataset. Its technical approach is mainly divided into two categories: methods based on generative adversarial networks and methods based on unsupervised clustering algorithms for model fine-tuning. The former uses generative adversarial networks to generate training data with various distributions by changing the background or style transfer of the target domain images, while the latter uses clustering algorithms to generate pseudo-labels, thus solving the unsupervised person re-identification problem in a supervised manner.

[0004] However, current unsupervised cross-domain person re-identification techniques can only identify objects that have already appeared in the source domain within the target domain, failing to recognize unknown objects not present in the source domain. This limits the application scope of this technique, and the presence of unknown objects in the target domain affects the model's accuracy in recognizing known objects. Furthermore, if videos or images captured by different cameras exhibit significant domain differences, existing techniques based on generative adversarial networks or unsupervised clustering algorithms may result in poor cross-domain person re-identification accuracy due to the generation of invalid augmented data samples or erroneous pseudo-labels. Therefore, researching unsupervised cross-domain open set person re-identification techniques to simultaneously identify both known and unknown objects is of great significance. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-domain open set person re-identification method and system based on unsupervised domain adaptation, so as to obtain a balanced and stable cross-domain person re-identification effect.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A cross-domain open set person re-identification method based on unsupervised domain adaptation includes:

[0008] Step 1: Preprocess the source domain person image and the target domain person image to obtain the preprocessed source domain person image and target domain person image, which will serve as the original input images for the following modules;

[0009] Step 2: Construct a cross-domain feature adversarial alignment module: The cross-domain feature adversarial alignment module includes a feature extractor G and two cross-domain discriminators F1 and F2. By aligning the adversarial source domain and target domain features, the domain differences between them are reduced, and the re-identification of known objects is achieved.

[0010] Step 3: Constructing an integrated binary classification bias module D: The integrated binary classification bias module consists of multiple binary classification bias modules. It integrates the output results of multiple binary classification bias modules to identify unknown objects in the target domain, realizes the division of known and unknown objects in the target domain, and reduces the negative impact of unknown objects on the re-identification accuracy of known objects;

[0011] Step 4: Input the source domain person image into the feature extractor G to obtain a new source domain person image representation. Simultaneously input the new source domain person image representation into the cross-domain discriminators F1 and F2 to obtain the classification results of F1 and F2 on the source domain image.

[0012] Step 5: Optimize the feature extractor G and cross-domain discriminators F1 and F2 using multi-class cross-entropy loss;

[0013] Step 6: Input the target domain person image into the feature extractor G to obtain a new target domain person image representation. Simultaneously input the new target domain person image representation into the cross-domain discriminators F1 and F2 to obtain the classification results on the target domain images of F1 and F2. Use the Manhattan distance to measure the difference between the two classification results.

[0014] Step 7: Calculate the entropy of the classification results of F1 and F2 to quantitatively measure the uncertainty of their classification on the target domain image. Minimize the entropy of their classification results to improve the confidence of the output. Repeat steps 4 and 5 to maximize the classification difference between F1 and F2 on the target domain image while ensuring that F1 and F2 classify correctly on the source domain image, so as to find samples distributed near the decision boundary in the target domain and optimize the cross-domain discriminators F1 and F2.

[0015] Step 8: Repeat step 6 to minimize the entropy of the F1 and F2 classification results to improve the confidence of the output, so that the decision boundary moves away from the low-density region of the sample distribution; at the same time, minimize the classification difference between the two on the target domain image to optimize the feature extractor G so that the extracted features are as far away from the decision boundary as possible, and obtain a more discriminative image representation.

[0016] Step 9: Repeat step 4, input the new source domain human image representation into the integrated binary classification bias module D, input it into the corresponding multiple binary classification bias modules according to the source domain image label, calculate the corresponding bias score as a quantitative measure of the degree of deviation between the image and the corresponding class, so that the bias scores of images of the same class are as close as possible to each other, while the bias scores of images of different classes are as different as possible to optimize the parameters of the integrated binary classification bias module D.

[0017] Repeat training steps four through nine until the model converges;

[0018] Step 10: After the model training is completed, in the model deployment phase, the target domain image is first input into the cross-domain feature adversarial alignment module to obtain the F1 classification result. Then, the new image representation extracted by the feature extractor G is input into the ensemble binary classification bias module D. Based on the F1 classification result, the bias score of this image is obtained. Based on the bias score of this image, it is determined whether it belongs to a known or unknown object. If it belongs to a known object, the F1 classification result is used as the final label of this image, realizing cross-domain open set person re-identification.

[0019] Furthermore, the preprocessing of the source and target domain images in step one includes image size normalization, random rotation, random cropping, and normalization operations to obtain preprocessed enhanced image data.

[0020] Furthermore, the backbone network of the feature extractor G in step two is a residual network ResNet-18 with the fully connected layers removed; the cross-domain discriminators F1 and F2 in step two include Dropout layers that can randomly mask neurons, fully connected layers, batch normalization layers, and nonlinear activation function layers.

[0021] Furthermore, the integrated binary classification bias module D in step three includes C. s There are two binary classification bias modules, where C s The number of known labels in the source domain image; the binary classification bias module contains three fully connected layers, where the last layer has 1 neuron and outputs a scalar value representing the bias score of the image.

[0022] Furthermore, the multi-class cross-entropy loss in step five is formulated as follows:

[0023]

[0024] Where, N s For source domain Size, x i For the preprocessed source domain image data, y i For its corresponding label, P s The distribution that the source domain image samples follow; and These are the outputs of cross-domain discriminators F1 and F2, respectively. When y i When =c, I(y) i c) = 1, when y i When ≠c, I(y) i c) = 0.

[0025] Furthermore, the formula for calculating the Manhattan distance in step six is ​​described as follows:

[0026]

[0027] Furthermore, the formulas for calculating the entropy of the classification results F1 and F2 in step seven are as follows:

[0028]

[0029] The optimization objectives for the cross-domain discriminators F1 and F2 in step seven are as follows:

[0030]

[0031] in, For the target domain, N t L represents the size of the target domain. adv (xj The calculation formula is as follows:

[0032]

[0033] in,

[0034] L e (x j The formula for calculating ) is:

[0035]

[0036] Furthermore, the optimization objective of the feature extractor G in step eight is:

[0037]

[0038] Furthermore, the formula for calculating the deviation score in step nine is described as follows:

[0039]

[0040] Among them, D c For the c-th binary classification bias module in D, the optimization objective of the binary classification bias module D in step nine is:

[0041]

[0042] Among them, L D This is the deviation loss.

[0043] Furthermore, the process of determining whether an image belongs to a known or unknown object based on its deviation score in step ten is as follows: First, the target domain image sample x... j The values ​​∈T are sequentially input into the model to obtain the F1 classification result. And each binary classification bias module D in D c Corresponding deviation score Calculate D c The mean μ of the deviation scores of all corresponding samples c and standard deviation σ c The calculation formula is described as follows:

[0044]

[0045]

[0046] According to the Cantali inequality:

[0047]

[0048] Taking ε=ασ, α>0, we get:

[0049]

[0050] Calculate D using the Cantali inequality. c The corresponding deviation score threshold r c The calculation formula is described as follows:

[0051] r c =μ c +ασ c

[0052] Finally, based on the F1 classification results Get x j Corresponding deviation score if Then determine x j For unknown objects; if Then x j It will be classified as the category labeled κ.

[0053] A cross-domain open set person re-identification system based on unsupervised domain adaptation is characterized by comprising five parts: a preprocessing module, a cross-domain feature adversarial alignment module, an integrated binary classification bias module, a training optimization module, and a deployment module;

[0054] The preprocessing module is used to perform preprocessing on the source and target domain images, such as random rotation, random cropping, and normalization, to obtain the preprocessed image.

[0055] The cross-domain feature adversarial alignment module is used to align the feature distributions of source and target domain samples in an alternating adversarial training manner, so as to identify objects that have appeared in the source domain in the target domain;

[0056] The integrated binary classification bias module is used to identify unknown objects in the target domain that have not appeared in the source domain, thereby separating known and unknown objects in the target domain and reducing the negative impact of unknown objects on the re-identification accuracy of known objects.

[0057] The training optimization module consists of three parts: the first part is the feature extraction optimization part, which optimizes the feature extractor by combining multi-class cross-entropy loss, difference loss and bias loss; the second part is the cross-domain discriminator optimization part, which optimizes the cross-domain discriminator by combining difference loss and entropy minimization loss; the third part is the integrated binary classification bias module optimization part, which optimizes the integrated binary classification bias module separately by using bias loss.

[0058] The deployment module is used to calculate the threshold of the deviation score corresponding to each binary classification deviation module in the integrated binary classification deviation module, which is used to determine whether the image belongs to a certain known class in the source domain.

[0059] This invention is a method for cross-domain person re-identification, and compared with existing technologies, this invention has the following advantages:

[0060] (1) In view of the shortcomings of the current cross-domain person re-identification algorithm, which cannot identify unknown objects that have not appeared in the source domain, this invention proposes a cross-domain open set person re-identification algorithm, which can not only re-identify known objects that have appeared in the source domain in the target domain, but also identify unknown objects that have not appeared in the source domain in the target domain.

[0061] (2) This invention does not adopt the technical route based on generative adversarial networks and unsupervised clustering algorithms, thus avoiding the phenomenon of generating invalid augmented data samples and incorrect pseudo-labels due to excessive domain differences, resulting in poor re-identification accuracy. It proposes a cross-domain adversarial alignment module and an integrated binary classification bias module, and adopts a joint training method to simultaneously perform cross-domain feature distribution adversarial alignment and unknown object identification, achieving a balanced and robust re-identification effect. Balanced means that while ensuring efficient identification of known objects, it can also identify unknown objects with high accuracy. Robust means that it can maintain stable identification accuracy when facing different proportions of unknown objects in the source domain. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart of a cross-domain open set person re-identification method based on unsupervised domain adaptation provided in an embodiment of the present invention;

[0064] Figure 2 This is a framework diagram of a cross-domain open set person re-identification method based on unsupervised domain adaptation provided in an embodiment of the present invention; Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The purpose of this invention is to provide a cross-domain open set person re-identification method based on unsupervised domain adaptation, so as to obtain a balanced and stable cross-domain person re-identification effect.

[0067] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] like Figure 1 As shown, cross-domain open set person re-identification methods based on unsupervised domain adaptation include:

[0069] Step 101: Preprocess the source domain person image and the target domain person image to obtain preprocessed source domain person images and target domain person images. Specifically, this includes:

[0070] The source and target domain character images are subjected to image size normalization, random rotation, random cropping, and normalization operations to obtain preprocessed character images. The image size is normalized to 256*256 (number of horizontal pixels * number of vertical pixels). The random rotation and random cropping operations utilize image operation functions provided by the torchvision framework. The mean and standard deviation of the normalization operations used on the three image channels are μ = [0.485, 0.456, 0.406] and σ = [0.229, 0.224, 0.225], respectively.

[0071] Step 102: Constructing a cross-domain feature adversarial alignment module: The cross-domain feature adversarial alignment module includes a feature extractor G and two cross-domain discriminators F1 and F2; the feature extractor G is a ResNet-18 residual network with the fully connected layers removed; the cross-domain discriminators F1 and F2 have the same structure, both containing a Dropout layer that can randomly mask neurons, a fully connected layer, a batch normalization layer, and a non-linear activation function layer; wherein the masking coefficient of the Dropout layer is 0.5, and the non-linear activation function layer uses ReLU as the activation function; as shown Figure 2 The overall structure of the cross-domain feature adversarial alignment module is visible.

[0072] Step 103: Construct the integrated binary classification bias module D; D includes C s A binary classification bias module D c c∈{1,...,C} s}, where C s The number of known labels for the source domain image; the binary classification bias module D c It contains three fully connected layers, with the last layer having only one neuron, and outputs a scalar value representing the image's bias score. For example... Figure 2 The structure of the integrated binary classification bias module D is visible.

[0073] Step 104: Transfer the source domain image x i The input feature extractor G is used to obtain a new source domain character image representation z. i =G(x)i The new source domain image representation zi is simultaneously input into cross-domain discriminators F1 and F2 to obtain the classification results of F1 and F2 on the source domain image. Know

[0074] Step 105: Optimize the feature extractor G and cross-domain discriminators F1 and F2 using multi-class cross-entropy loss; the calculation formula for the multi-class cross-entropy loss is described as follows:

[0075]

[0076] Where, N s For source domain Size, x i For the preprocessed source domain image data, y i For its corresponding label, P s The distribution that the source domain image samples follow; and These are the outputs of cross-domain discriminators F1 and F2, respectively. When y i When =c, I(y) i c) = 1, when y i When ≠c, I(y) i c) = 0.

[0077] Step 106: Extract the target domain image of the person x j ∈T is input into the feature extractor G, and a new target domain person image representation z is obtained. j =G(x) j The new target domain character image representation z j The results are simultaneously input into cross-domain discriminators F1 and F2 to obtain the classification results on the target domain images of F1 and F2. and The Manhattan distance is used to measure the difference between the two classification results. The formula for calculating the Manhattan distance is described as follows:

[0078]

[0079] Step 107: Repeat steps 104 and 105 to calculate the entropy of the classification results of F1 and F2 to quantitatively measure the uncertainty of their classification on the target domain image. Minimize the entropy of their classification results to improve the confidence of the output. While ensuring that F1 and F2 classify correctly on the source domain image, maximize their classification difference on the target domain image to optimize the cross-domain discriminators F1 and F2. The formulas for calculating the entropy of the classification results of F1 and F2 are as follows:

[0080]

[0081] The optimization objectives of the optimized cross-domain discriminators F1 and F2 are:

[0082]

[0083] in, For the target domain, N t L represents the size of the target domain. adv (x j The calculation formula is:

[0084]

[0085] L e (x j The formula for calculating ) is:

[0086]

[0087] Step 108: Repeat step 106, minimizing the entropy of the F1 and F2 classification results to improve the output confidence, causing the decision boundary to move away from the low-density region of the sample distribution; simultaneously, minimizing the classification difference between the two on the target domain image to optimize the feature extractor G so that its extracted features are as far away from the decision boundary as possible, obtaining a more discriminative image representation; the optimization objective of the feature extractor G is:

[0088]

[0089] Step 109: Repeat step 104 to represent the new source domain person image z. i The input is fed into the integrated binary classification bias module D. Based on the source domain image label, it is input into multiple corresponding binary classification bias modules. The corresponding bias score is calculated as a quantitative measure of the degree of deviation between the image and the corresponding class. The goal is to optimize the parameters of the integrated binary classification bias module D by making the bias scores of images of the same class as close as possible to each other while making the bias scores of images of different classes as different as possible. The formula for calculating the bias score is described as follows:

[0090]

[0091] Among them, D c Let c be the c-th binary classification bias module in D; the optimization objective of the binary classification bias module D is:

[0092]

[0093] Among them, L D The bias loss is used to optimize the bias loss. The objective of this optimization is to make the bias scores of images of the same class as close as possible to the hyperparameter c in the binary classification bias module, while the bias scores of images of different classes should differ from c by at least a size of m.

[0094] Repeat steps 104 to 109 until the model converges.

[0095] Step 110: After model training is complete, in the model deployment phase, first, the target domain image x... j The values ​​∈T are sequentially input into the cross-domain feature adversarial alignment module to obtain the F1 classification result. The new image representation z extracted by the feature extractor G j =G(x) j ); will z j The input is fed into the integrated binary classification bias module D, resulting in each binary classification bias module D in D. c Corresponding deviation score Calculate D c The mean μ of the deviation scores of all corresponding samples c and standard deviation σ c The calculation formula is described as follows:

[0096]

[0097]

[0098] Calculate D using the Cantali inequality. c The corresponding deviation score threshold r c The calculation formula is described as follows:

[0099] r c =μ c +ασ c

[0100] Finally, based on the F1 classification results Get x j Corresponding deviation score if Then determine x j For unknown objects; if Then x j It will be classified as the category labeled κ.

[0101] This embodiment also provides a cross-domain open set person re-identification system based on unsupervised domain adaptation, characterized by comprising five parts: a preprocessing module, a cross-domain feature adversarial alignment module, an integrated binary classification bias module, a training optimization module, and a deployment module;

[0102] The preprocessing module is used to perform preprocessing on the source and target domain images, such as random rotation, random cropping, and normalization, to obtain the preprocessed image.

[0103] The cross-domain feature adversarial alignment module is used to align the feature distributions of source and target domain samples in an alternating adversarial training manner, so as to identify objects that have appeared in the source domain in the target domain;

[0104] The integrated binary classification bias module is used to identify unknown objects in the target domain that have not appeared in the source domain, thereby separating known and unknown objects in the target domain and reducing the negative impact of unknown objects on the re-identification accuracy of known objects.

[0105] The training optimization module consists of three parts: the first part is the feature extraction optimization part, which optimizes the feature extractor by combining multi-class cross-entropy loss, difference loss and bias loss; the second part is the cross-domain discriminator optimization part, which optimizes the cross-domain discriminator by combining difference loss and entropy minimization loss; the third part is the integrated binary classification bias module optimization part, which optimizes the integrated binary classification bias module separately by using bias loss.

[0106] The deployment module is used to calculate the threshold of the deviation score corresponding to each binary classification deviation module in the integrated binary classification deviation module, which is used to determine whether the image belongs to a known class or an unknown class in the source domain.

[0107] This embodiment can also be implemented using the following code:

[0108]

[0109]

[0110] This invention provides a method and system for cross-domain open set person re-identification based on unsupervised domain adaptation. Addressing the problem that current cross-domain person re-identification methods and systems can only re-identify known objects and cannot identify unknown objects, this invention achieves simultaneous identification of both known and unknown objects by constructing a cross-domain adversarial alignment module and an integrated binary classification bias module. This method does not employ techniques based on generative adversarial networks or unsupervised clustering algorithms, thus avoiding the phenomenon of these methods generating invalid augmented data samples and erroneous pseudo-labels due to excessive domain differences, resulting in poor re-identification accuracy. It achieves a balanced and robust re-identification effect; balanced meaning it ensures efficient identification of known objects while also identifying unknown objects with high accuracy; robust meaning it maintains stable identification accuracy even when faced with different proportions of unknown objects in the source domain.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0112] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for cross-domain open set person re-identification based on unsupervised domain adaptation, characterized in that, include: Step 1: Preprocess the source domain image and the target domain image to obtain the preprocessed source domain image and target domain image; Step 2: Construct a cross-domain feature adversarial alignment module; the cross-domain feature adversarial alignment module includes a feature extractor G and two cross-domain discriminators F1 and F2; Step 3: Construct an integrated binary classification bias module D; the integrated binary classification bias module consists of multiple binary classification bias modules; Step 4: Input the source domain person image into the feature extractor G to obtain a new source domain person image representation. Simultaneously input the new source domain person image representation into the cross-domain discriminators F1 and F2 to obtain the classification results of F1 and F2 on the source domain image. Step 5: Optimize the feature extractor G and cross-domain discriminators F1 and F2 using multi-class cross-entropy loss; Step 6: Input the target domain person image into the feature extractor G to obtain a new target domain person image representation. Simultaneously input the new target domain person image representation into the cross-domain discriminators F1 and F2 to obtain the classification results on the target domain images of F1 and F2. Use the Manhattan distance to measure the difference between the two classification results. Step 7: Calculate the entropy of the classification results of F1 and F2; repeat steps 4 and 5; while ensuring that F1 and F2 classify correctly on the source domain image, maximize the classification difference between the two on the target domain image to optimize the cross-domain discriminators F1 and F2. Step 8: Repeat Step 6, and minimize the entropy of the F1 and F2 classification results to improve the confidence of the output; at the same time, minimize the classification difference between the two on the target domain image to optimize the feature extractor G; Step 9: Repeat step 4 to calculate the deviation score corresponding to the source domain image, so that the deviation scores of images of the same class are as close as possible to each other, while the deviation scores of images of different classes are greater than the preset threshold to optimize the parameters of the integrated binary classification deviation module D. Repeat training steps four through nine until the model converges; Step 10: After the model training is completed, in the model deployment phase, the target domain image is input into the model to obtain the corresponding bias score of the image. Based on the bias score of the image, it is determined whether it belongs to a known or unknown object.

2. The cross-domain open set person re-identification method based on unsupervised domain adaptation according to claim 1, characterized in that, In step two, the backbone network of the feature extractor G is a ResNet-18 residual network with the fully connected layers removed; the cross-domain discriminators F1 and F2 in step two include Dropout layers that can randomly mask neurons, fully connected layers, batch normalization layers, and non-linear activation function layers. The integrated binary classification bias module D in step three includes: A binary classification bias module, in which The number of labels in the known source domain image; the binary classification bias module contains three fully connected layers, where the last layer has 1 neuron and outputs a scalar value representing the bias score of the image.

3. The method for cross-domain open set person re-identification based on unsupervised domain adaptation according to claim 1, characterized in that, The formula for calculating the multi-class cross-entropy loss in step five is as follows: Where, N s For source domain Size, x i For the preprocessed source domain image data, y i For its corresponding tag, The distribution that the source domain image samples follow; and These are the outputs of cross-domain discriminators F1 and F2, respectively. When y i When = c, When y i When ≠c, 4. The method for cross-domain open set person re-identification based on unsupervised domain adaptation according to claim 1, characterized in that, The optimization objectives for the cross-domain discriminators F1 and F2 in step seven are as follows: in, For the target domain, N t The size of the target domain; The calculation formula is as follows: in, Distance to Manhattan; The calculation formula is as follows: in, and Classification results for F1 and F2 and Entropy; The optimization objective of the feature extractor G in step eight is: Among them, D c Let D be the c-th binary classification bias module; the optimization objective of the binary classification bias module D in step nine is: in, For deviation loss, For x i The corresponding deviation score; The process of determining whether an image belongs to a known or unknown object based on its deviation score in step ten is as follows: First, the target domain image samples... The data is input into the model sequentially to obtain the F1 classification result. And each binary classification bias module D in D c Corresponding deviation score Calculate D c The mean μ of the deviation scores of all corresponding samples c and standard deviation σ c ; Calculate D using the Cantali inequality. c The corresponding deviation score threshold r c The calculation formula is described as follows: r c =μ c +as c Finally, based on the F1 classification results Get x j Corresponding deviation score if Then determine x j If it is an unknown object; otherwise, then x j It will be categorized as the one labeled κ.

5. A cross-domain open set person re-identification system based on unsupervised domain adaptation, characterized in that, It consists of five parts: a preprocessing module, a cross-domain feature adversarial alignment module, an integrated binary classification bias module, a training optimization module, and a deployment module; The preprocessing module is used to perform random rotation, random cropping, and normalization preprocessing on the source and target domain images to obtain the preprocessed image. The cross-domain feature adversarial alignment module is used to align the feature distributions of source and target domain samples in an alternating adversarial training manner, so as to identify objects that have appeared in the source domain in the target domain; The integrated binary classification bias module is used to identify unknown objects in the target domain that have not appeared in the source domain, thereby separating known and unknown objects in the target domain and reducing the negative impact of unknown objects on the re-identification accuracy of known objects. The training optimization module consists of three parts; the first part is the feature extraction optimization part, which uses multi-class cross-entropy loss, difference loss and bias loss to comprehensively optimize the feature extractor. The second part is the cross-domain discriminator optimization, which uses difference loss and entropy minimization loss to comprehensively optimize the cross-domain discriminator; the third part is the integrated binary classification bias module optimization, which uses bias loss to separately optimize the integrated binary classification bias module. The deployment module is used to calculate the threshold of the deviation score corresponding to each binary classification deviation module in the integrated binary classification deviation module, which is used to determine whether the image belongs to a known class or an unknown class.

Images