A passive data unsupervised domain adaptation method for semantic segmentation
By constructing a virtual source domain model and an uncertainty-aware self-training method, the problems of privacy and data transmission difficulties in unsupervised domain adaptation are solved. This achieves high-performance semantic segmentation without accessing source domain data, protecting user privacy and reducing the need for labeled data.
Patent Information
- Application Number
- CN202211435678.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-16
- Publication Date
- 2026-02-06
- Estimated Expiration
- 2042-11-16
AI Technical Summary
Existing unsupervised domain adaptation methods require access to the original source domain data, which leads to privacy leaks and difficulties in data transmission, and they rely on large-scale datasets with manual annotations.
A virtual source domain model is constructed, which is initialized using a pre-trained source domain prediction model. Through weighted entropy loss and regularization optimization, combined with uncertainty-aware self-training and prior distribution-aware adversarial training methods, semantic segmentation without accessing the original source domain data is achieved.
This approach improves the performance of semantic segmentation models, protects user privacy, and reduces reliance on large-scale labeled data without accessing source domain data.
Smart Images
Figure CN115690428B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of unsupervised domain adaptation, and more particularly to a passive data unsupervised domain adaptation method for semantic segmentation. BACKGROUND
[0002] The purpose of semantic segmentation is to assign a semantic label to each pixel in an image. In recent years, segmentation models based on convolutional neural networks have made significant progress and have been widely used in computer vision fields such as autonomous driving, robotics, and disease diagnosis. However, the remarkable performance of current semantic segmentation methods usually relies on large-scale pixel-level datasets with manual annotations. To avoid the time-consuming and laborious manual annotation, many unsupervised domain adaptation (UDA) methods have been proposed to use a model trained on computer-generated annotated images for unannotated real images. Unsupervised domain adaptation (UDA) for semantic segmentation can transfer knowledge (with low-cost annotations) obtained from synthetic data (source domain) to real images (target domain). All current UDA methods explicitly align the distribution of the target domain with that of the source domain for feasible adaptation. Therefore, these methods are usually based on the assumption of access to source domain labeled data, which limits their applications in practical situations where there are privacy, storage, or transmission issues. For example, users or companies do not want their data to be leaked to others, and some source domain data is too large to be quickly transmitted to the cloud.
[0003] Therefore, how to provide an unsupervised domain adaptation method with high security performance without accessing the original source domain data is a problem that those skilled in the art need to solve. SUMMARY
[0004] Therefore, the present application provides an unsupervised domain adaptation method for semantic segmentation without accessing the original source domain data, only requiring a pre-trained source domain prediction model and unannotated target domain data, which can protect user privacy.
[0005] To achieve the above purpose, the present application adopts the following technical solutions:
[0006] An unsupervised domain adaptation method for semantic segmentation, comprising the following steps:
[0007] A virtual source domain construction model is constructed, and the virtual source domain construction model is initialized using a pre-trained source domain prediction model;
[0008] The virtual source domain construction model is optimized using a weighted entropy loss and a regularization term;
[0009] The target domain samples are predicted by using the optimized virtual source domain construction model, and the virtual source domain and the uncertain target domain are divided according to the prediction confidence;
[0010] A semantic segmentation model is constructed, and the semantic segmentation model is initialized by using a pre-trained source domain prediction model;
[0011] The semantic segmentation model is optimized by using an uncertainty-aware self-training method and a prior distribution-aware adversarial training method;
[0012] The distribution between the uncertain target domain and the virtual source domain is aligned based on the optimized semantic segmentation model.
[0013] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the source domain prediction model is trained and converged by using data of a real source domain; and the parameters of the virtual source domain construction model are obtained by using the parameters of the trained source domain prediction model.
[0014] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, a weighted entropy loss is used to minimize the entropy of the target samples output by the virtual source domain output model and higher than a preset confidence, and the entropy of the target samples output by the virtual source domain output model and lower than the preset confidence is kept unchanged; and the virtual source domain construction model is optimized by using the following formula:
[0015]
[0016] wherein, P t = softmax(φ V (X t ));
[0017] In the above formula, represents a weighted entropy loss; E represents an expectation; exp represents an exponential operation; m represents a hyperparameter; X t represents a target domain picture; H(X t )∈R H×W represents a normalized pixel-level entropy map; ⊙ represents pixel-by-pixel multiplication; exp(-mH(X t )) represents a nonlinear weighting function; C represents the number of categories of a data set; φ V represents a virtual source domain construction model; φ V (X t ) represents the output of X t after being input into the virtual source domain construction model φ V ; P t represents a prediction probability output by the virtual source domain construction model; h, w, and c represent the size of P t .
[0018] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the consistency regularization loss between the prediction of the source domain prediction model and the prediction of the virtual domain construction model is calculated by the following formula:
[0019]
[0020] wherein, I(*) represents an indicator function, I(*) = 1 when * ≥ 0.9, otherwise I(*) = 0; represents the source domain prediction model.
[0021] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the partitioning method of the virtual source domain and the uncertain target domain is as follows:
[0022] The average entropy of a certain target domain picture in the target domain sample is calculated by the optimized virtual source domain construction model prediction;
[0023] The average entropies of all target domain pictures in the target domain sample are arranged in ascending order, and the top 66% of the target domain pictures are taken as the virtual source domain, and the remaining 34% of the target domain pictures are taken as the uncertain target domain.
[0024] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the loss of the optimized virtual source domain construction model is calculated by the following formula to represent the distribution difference between the virtual source domain and the real source domain:
[0025]
[0026] wherein, is the consistency regularization loss, is the weighted entropy loss, λ re and λ went are the weights of the consistency regularization loss and the weighted entropy loss, respectively.
[0027] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the uncertainty-aware self-training method comprises the following steps:
[0028] A set of virtual source domain construction models with dropout are defined
[0029] For each picture X d in the virtual source domain and the uncertain target domain, perform N times of forward propagation to obtain a set of prediction outputs wherein, d ∈ {v, u},
[0030] A pseudo label in the predicted output is obtained using the following formula:
[0031]
[0032] In the above formula, represents the generated pseudo label, and argmax is a specific function;
[0033] The obtained pseudo label is selected using the following formula:
[0034]
[0035] B(X d represents the value of uncertainty;
[0036] The loss of the uncertainty self-training stage is determined using the following formula:
[0037]
[0038] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation using passive data, the prior distribution-aware adversarial training method improves the object class mismatch, class imbalance, and predicted uncertainty between the virtual source domain and the uncertain target domain during the alignment process between the virtual source domain and the uncertain target domain; wherein the process of improving the object class mismatch between the virtual source domain and the uncertain target domain is as follows:
[0039] The class probabilities of the virtual source domain and the uncertain target domain are calculated using the following formula respectively and
[0040] wherein,
[0041] A threshold θ is introduced to select the domain-shared classes in the virtual source domain and the uncertain target domain, and to exclude the domain-specific classes; the selection formula is as follows:
[0042]
[0043] The above formula indicates that represents whether the current class is selected, represents whether the current class is selected for alignment or excluded, if or is less than θ, then is 0, otherwise is 1; c is the class;
[0044] The class imbalance between the virtual source domain and the uncertain target domain is improved using the class weight, and the calculation formula of the class weight is as follows:
[0045]
[0046] wherein, represents the class-balanced weight value of the virtual source domain; represents the class-balanced weight value of the uncertain target domain;
[0047] The uncertainty weight is used to improve the uncertainty factors of the predictions of the virtual source domain and the uncertain target domain, and the uncertainty weight The calculation formula is as follows:
[0048]
[0049] wherein, w[c * ] represents the value of the index c * in the w matrix; Xone represents a matrix full of 1; I d represents the entropy value of the picture; If the uncertainty is high, the weight is low, and if the uncertainty is low, the weight is high. represents the weight matrix of the uncertainty.
[0050] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the loss calculation method of the prior distribution perception adversarial training method stage is as follows:
[0051]
[0052] wherein, and respectively represent the entropy value of the virtual source domain picture and the entropy value of the uncertain target domain.
[0053] Preferably, in the above-mentioned unsupervised domain adaptation method for semantic segmentation without source data, the loss of the optimized semantic segmentation model is calculated by the following formula, which is used to represent the data uncertainty and the prior distribution difference between the virtual source domain and the real source domain;
[0054]
[0055] wherein, λ u_self = 1, λ wadυ = 0.01, respectively represent the weight of the uncertainty perception self-training stage and the prior distribution perception adversarial training stage.
[0056] Compared with the prior art, the technical scheme can provide a passive data unsupervised domain self-adaptive method for semantic segmentation, which firstly constructs a set of source domain virtual data to simulate source domain distribution by identifying target domain high-confidence samples predicted by a pre-trained source domain prediction model. Then, by analyzing data attributes in a cross-domain semantic segmentation task, an uncertainty-aware self-training method and a prior distribution-aware adversarial training method are used to align the virtual source domain and the target domain with self-training and adversarial learning strategies. The method can directly learn the target domain distribution from noisy pseudo labels, and can improve object class mismatch, class imbalance and prediction uncertainty between different domains. BRIEF DESCRIPTION OF DRAWINGS
[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.
[0058] Figure 1 The drawings are flowcharts of the passive data unsupervised domain self-adaptive method for semantic segmentation provided by the present application. DETAILED DESCRIPTION
[0059] The technical solutions in the embodiments of the present application will be described clearly and completely in the embodiments of the present application in combination with the drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0060] As shown in Figure 1 The present application discloses a passive data unsupervised domain self-adaptive method for semantic segmentation, comprising the following steps:
[0061] A virtual source domain construction model is constructed, and the virtual source domain construction model is initialized by using a pre-trained source domain prediction model;
[0062] The virtual source domain construction model is optimized by using a weighted entropy loss and a regularization term;
[0063] The target domain samples are predicted by using the optimized virtual source domain construction model, and the virtual source domain and the uncertain target domain are divided according to the prediction confidence;
[0064] A semantic segmentation model is constructed, and the semantic segmentation model is initialized by using a pre-trained source domain prediction model;
[0065] The semantic segmentation model is optimized by using an uncertainty perception self-training method and a prior distribution perception adversarial training method;
[0066] The distribution between the uncertain target domain and the virtual source domain is aligned based on the optimized semantic segmentation model.
[0067] The algorithm comprises two stages, namely 1) virtual source domain construction, and 2) uncertainty and prior distribution perception domain adaptation.
[0068] In the first stage, it is suggested to construct a virtual source domain by selecting highly confident target domain samples to imitate the source domain distribution; then the inter-domain gap between the virtual source domain and the real source domain is further reduced by the proposed weighted entropy minimization method; the gap refers to the distribution gap between the virtual source domain and the real source domain.
[0069] High confidence refers to the prediction probability of the model, when the prediction probability of the model for the sample is high, we simply refer to the high confidence of the sample, and the model is confident for the prediction made by itself. High confidence means high confidence.
[0070] The weighted entropy minimization method refers to minimizing the weighted entropy loss and consistency regularization loss.
[0071] In the second stage, by considering data uncertainty and prior distribution difference, a carefully designed weighting mechanism is proposed for distribution alignment between the target domain and the virtual source domain. In addition, a new uncertainty perception self-training method is proposed, which can directly learn the target domain distribution from noisy pseudo labels by considering model uncertainty.
[0072] Next, each step described above will be further described.
[0073] 1. Construction of virtual source domain
[0074] The source domain prediction model is a model pre-trained and converged using source domain data, trained using source domain images and labels. The source domain prediction model remains unchanged after training and convergence, and is defined as the source domain pre-training model Initialization can be understood as parameter copying, using the source domain prediction model Initialize the virtual source domain construction model φ V is the parameter of copied to V .
[0075] 1) Weighted entropy loss: entropy minimization optimization is performed on low entropy (i.e. high confidence) target samples, and high entropy samples with low prediction confidence are kept unchanged.
[0076] The weighted entropy loss is used to minimize the entropy of the target sample output by the virtual source domain output model and higher than the preset confidence, and the entropy value of the target sample output by the virtual source domain output model and lower than the preset confidence is kept unchanged; the virtual source domain construction model is optimized by weighted entropy minimization using the following formula:
[0077]
[0078] wherein, P t =softmax(φ V (X t ));
[0079] In the above formula, denotes the weighted entropy loss; E denotes expectation; exp denotes exponential operation; m denotes a hyperparameter; X t denotes a target domain picture; H(X t )∈R H×W denotes a normalized pixel-level entropy map; ⊙ denotes pixel-by-pixel multiplication; exp(-mH(X t )) denotes a nonlinear weighting function; C denotes the number of categories of a data set; φ V denotes a virtual source domain construction model; φ V (X t ) denotes the output of X t after being sent to the virtual source domain construction model φ V ; P t denotes the predicted probability output by the virtual source domain construction model; h, w, and c denote the size of P t .
[0080] 2) Consistency regularization loss
[0081] In order to avoid the propagation of errors caused by overconfident predictions, the present application proposes a consistency regularization loss between the prediction of the original source domain prediction model and the prediction of the virtual domain construction model. The calculation formula of the consistency regularization loss is as follows:
[0082]
[0083] wherein, I(*) denotes an indicator function, I(*)=1 when *≥0.9, otherwise I(*)=0; denotes a source domain prediction model.
[0084] The virtual source domain construction model φ V is trained using formula (1) and formula (2), and the virtual domain construction model φ V is used to construct a virtual source domain based on the prediction of unlabelled target samples.
[0085] For one picture in the target domain sample, the average entropy of the whole picture is calculated using the prediction of V Then the average entropy is sorted from small to large on the whole dataset, and the top 66% is selected as the virtual source domain The remaining 34% is the uncertain target domain
[0086] The loss of the optimized virtual source domain construction model is calculated by the following formula, which represents the distribution difference between the virtual source domain and the real source domain.
[0087]
[0088] Where, is the consistency regularization loss, is the weighted entropy loss, and λ re and λ went are the weights of the consistency regularization loss and the weighted entropy loss, respectively; formula (3) is the overall loss of the first stage, which is the sum of the above two losses, λ re and λ went are set to 0.01 and 1, respectively.
[0089] 2. Uncertainty and prior distribution aware domain adaptation
[0090] Next, the invention uses a weighted adversarial loss to indirectly align the distribution between the uncertain target domain and the virtual source domain, and uses an uncertainty-aware self-training method to directly learn the target domain distribution from noisy pseudo labels.
[0091] 1) Uncertainty-aware self-training
[0092] The commonly used self-training method in unsupervised domain adaptation is to manually set a threshold according to the prediction confidence, but there are still some unresolved problems: (1) If the threshold is set too high, the contribution of pseudo-labeled samples to performance improvement is not great; (2) If the threshold is set too low, it may have an adverse effect on the model due to the noise of the pseudo label; (3) A single threshold may not be suitable for all classes and datasets, which is another obstacle to effective sample selection.
[0093] Some recent studies have proposed Bayesian neural networks to capture the uncertainty of the model and have successfully applied them to various computer vision tasks. Bayesian neural networks replace deterministic network weight parameters with distributions over these parameters and average over all possible weights to capture the uncertainty of the model. To this end, the invention uses the advances in Bayesian deep learning to obtain reliable prediction uncertainty of the network and perform effective self-training.
[0094] Dropout is used to approximate inference in Bayesian neural networks. First, define a set of dummy source domains with dropout to build the model For each image X d ,d∈{v,u}, Perform N forward propagation to obtain a set of predictions The pseudo-labels are obtained as follows:
[0095]
[0096] In the above formula, represents the generated pseudo-label, and argmax is a specific function;
[0097] Select the pseudo-label obtained by using the following formula:
[0098]
[0099] B(X d ) represents the value of uncertainty;
[0100] The loss of the uncertainty self-training stage is determined by using the following formula:
[0101]
[0102] 2) Uncertainty prior distribution-aware adversarial training
[0103] There are some specific characteristics (such as object class mismatch between different domains, class imbalance, prediction uncertainty, etc.) in cross-domain semantic segmentation that need to be further considered. Based on these characteristics, the present application proposes a weighted adversarial learning method for domain alignment. The overall weighted adversarial target is defined as:
[0104]
[0105] Where two weighted parameters {β v ,β u} are defined and learned from the model output of each mini-batch.
[0106] Specifically, the present application embodiment takes batch size 2 as an example for illustration, one image X v from the dummy source domain and one image X u from the uncertainty target domain are in a mini-batch. X v and X u are sent into the semantic segmentation model to generate two probability maps and
[0107] ① For the problem of different field categories not matching:
[0108] Since an image usually cannot contain all categories, the virtual source domain and the target domain in a batch of images contain both field-shared and field-specific categories. Therefore, it is a suboptimal choice to align all categories without any consideration. The process of the embodiment of the present application for improving the object category mismatch between the virtual source domain and the uncertain target domain is as follows:
[0109] The category probability of the virtual source domain and the uncertain target domain is calculated respectively by using the following formula and
[0110] wherein,
[0111] The threshold θ is introduced to select the field-shared categories in the virtual source domain and the uncertain target domain, and to exclude the field-specific categories; the selection formula is as follows:
[0112]
[0113] The above formula indicates that, represents whether the current category is selected, represents whether the current category is selected for alignment or excluded, if or is less than θ, then is 0, otherwise is 1; c is the category.
[0114] ② The problem of category imbalance:
[0115] The category imbalance between the virtual source domain and the uncertain target domain is improved by using the category weight, and the calculation formula of the category weight is as follows:
[0116]
[0117] wherein, represents the category balance weight value of the virtual source domain; represents the category balance weight value of the uncertain target domain;
[0118] ③ The problem of prediction uncertainty
[0119] The prediction uncertainty of the virtual source domain and the uncertain target domain is improved by using the uncertainty weight, and the calculation formula of the uncertainty weight is as follows:
[0120]
[0121] wherein,
[0122] w[c * ] represents the value of the index c * in the w matrix; Xone represents a matrix full of 1; I d represents the entropy value of the picture; represents that the weight is low when the uncertainty is high, and the weight is high when the uncertainty is low; represents the weight matrix of the uncertainty.
[0123] The loss of the optimized semantic segmentation model is calculated by the following formula, to represent the data uncertainty and prior distribution difference between the virtual source domain and the real source domain;
[0124]
[0125] wherein, λ u_self = 1, λ wadυ = 0.01, respectively representing the weight of the uncertainty-aware self-training stage and the prior distribution-aware adversarial training stage.
[0126] As shown in Tables 2(a)-(b), the differences between the unsupervised domain adaptation method proposed in the present application and the latest unsupervised domain adaptation method in three cross-domain semantic segmentation tasks are compared.
[0127] Table 2(a)
[0128]
[0129]
[0130] Table 2(b)
[0131]
[0132] It can be observed that the method of the present application (even without source domain data) is 1.6% mIoU better than the most advanced unsupervised domain adaptation method (with source domain data) in the GTA5 to Cityscapes task, and 1.8% mIoU better in the SYNTHIA to Cityscapes task. From Synscapes and Cityscapes, our method gets mIoU of 54.5%, still 0.3% better in performance than the most advanced method.
[0133] The various embodiments described in this specification are implemented in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.
[0134] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A passive data-based unsupervised domain adaptive method for semantic segmentation, characterized in that, Includes the following steps: A virtual source domain construction model is constructed, and the virtual source domain construction model is initialized using a pre-trained source domain prediction model; the source domain prediction model is trained and converged using data from the real source domain; the parameters of the virtual source domain construction model are the parameters of the pre-trained source domain prediction model. The virtual source domain model is optimized using weighted entropy loss and regularization terms; The optimized virtual source domain is used to build a model to predict the target domain samples, and the virtual source domain and the uncertain target domain are divided according to the prediction confidence. Construct a semantic segmentation model and initialize the semantic segmentation model using a pre-trained source domain prediction model; An uncertainty-aware self-training method and a prior distribution-aware adversarial training method are used to optimize the semantic segmentation model; Align the distribution between the uncertain target domain and the virtual source domain based on the optimized semantic segmentation model; The uncertainty-aware self-training method includes the following steps: Define a set of virtual source domain construction models with dropout. For each image X in the virtual source domain and the uncertain target domain d Perform N forward propagations to obtain a set of predicted outputs. Where, d∈{v,u}, The pseudo-labels in the predicted output are obtained using the following formula; In the above formula, This represents the generated pseudo-labels, where argmax is a specific function. Use the following formula to select the obtained pseudo-labels; B(X d () indicates an uncertain value; The following formula is used to determine the loss during the uncertainty self-training phase; The prior distribution-aware adversarial training method improves the object class mismatch, class imbalance, and prediction uncertainty factors between the virtual source domain and the uncertain target domain during the process of aligning the distributions between them. Specifically, the process of improving the object class mismatch factors between the virtual source domain and the uncertain target domain is as follows: The following formulas are used to calculate the class probabilities of the virtual source domain and the uncertain target domain, respectively. and in, A threshold θ is introduced to select domain-shared categories in the virtual source domain and the uncertain target domain, while excluding domain-specific categories; the selection formula is as follows: The above formula means that, Indicates whether the current category is selected. Indicates whether the current category is selected or excluded. or If less than θ, then =0, otherwise 1 represents 'c'; 'c' represents the category. Class weights are used to improve the class imbalance between the virtual source domain and the uncertain target domain. The formula for calculating the class weights is as follows: in, This represents the class balance weight value of the virtual source domain; Represents the class balance weight value for the uncertain target domain; Uncertainty weights are used to improve the prediction of uncertainties in virtual source and uncertain target domains. The calculation formula is: in, w[c * ] indicates taking the index c in the w matrix. * The value of ; Xone represents a matrix consisting entirely of 1s; I d Represents the entropy value of the image; This indicates that higher uncertainty corresponds to lower weights, and lower uncertainty corresponds to higher weights. A weight matrix representing uncertainty.
2. The passive data unsupervised domain adaptive method for semantic segmentation according to claim 1, characterized in that, Weighted entropy loss is used to minimize the entropy of target samples output by the virtual source domain output model that have a higher confidence level than the preset level, while the entropy of target samples output by the virtual source domain output model that have a lower confidence level than the preset level remains unchanged. The following formula is used to minimize the weighted entropy of the virtual source domain model: in, P t =softmax(φ V (X t )); In the above formula, The weighted entropy loss is represented by E; the expected value is represented by exp; the hyperparameter is represented by m; and X represents the weighted entropy loss. t H(X) represents the target domain image; t )∈R H×W Represents a normalized pixel-level entropy map; ⊙ represents the pixel-by-pixel product; exp(-mH(X t )) represents a non-linear weighting function; C represents the number of classes in dataset l; φ V Represents the virtual source domain construction model; φ V (X t ) indicates that X t Send to virtual source domain to build model φ V The output after P; t The virtual source domain represents the predicted probability output by the model; h, w, c represent P. t The size.
3. The passive data unsupervised domain adaptive method for semantic segmentation according to claim 2, characterized in that, The consistency regularization loss between the predictions of the source domain prediction model and the predictions of the virtual domain construction model is calculated using the following formula: in, I(*) represents an indicator function. When *≥0.9, I(*)=1, otherwise I(*)=0; This represents the source domain prediction model.
4. The passive data unsupervised domain adaptive method for semantic segmentation according to claim 1, characterized in that, The method for dividing the virtual source domain and the uncertain target domain is as follows: The optimized virtual source domain is used to build a model to predict and calculate the average entropy of a target domain image in the target domain sample. Arrange the average entropy of all target domain images in the target domain sample from smallest to largest, take the top 66% of the target domain images as virtual source domains, and take the remaining 34% of the target domain images as uncertain target domains.
5. The passive data unsupervised domain adaptive method for semantic segmentation according to claim 3, characterized in that, The loss of the optimized virtual source domain model is calculated using the following formula to characterize the distribution difference between the virtual source domain and the real source domain; in, For consistency regularization loss, For weighted entropy loss, λ re and λ went These are the weights for consistency regularization loss and weighted entropy loss, respectively.
6. The passive data unsupervised domain adaptive method for semantic segmentation according to claim 1, characterized in that, The loss calculation method for the prior distribution-aware adversarial training phase is as follows: in, and These represent the entropy values of the virtual source domain image and the uncertain target domain, respectively.
7. The passive data unsupervised domain adaptive method for semantic segmentation according to claim 1, characterized in that, The loss of the optimized semantic segmentation model is calculated using the following formula to characterize the data uncertainty and prior distribution differences between the virtual source domain and the real source domain; Where, λ u-self =1,λ wadυ =0.01, representing the weights of the uncertainty-aware self-training stage and the prior distribution-aware adversarial training stage, respectively.