An open set image classification field self-adaption method based on self-paced learning

Abstract

The application discloses an open set image classification field self-adaption method based on self-step learning, first, the original image is preprocessed to obtain an image set, then a feature extraction module and a double multi-class classifier module are constructed and trained to align shared class features of source domain images and target domain images and separate target domain private class features, a multi-criteria cross-domain hybrid module is further constructed and trained, cross-domain hybrid images are generated by using the source domain images and the target domain images, and the shared class features are self-learned, and finally, a classification result of the target domain image is output. Compared with the existing open set image classification field self-adaption method, the application covers smooth and non-smooth class distribution, and does not need to empirically adjust the threshold for distinguishing common class images and private class images in the inference stage, so that the model has good robustness under different hyperparameters and experimental settings.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of open set image classification, and particularly relates to an open set image classification field adaptation method based on self-paced learning. BACKGROUND

[0002] Deep learning methods can achieve excellent performance on new data that is independent and identically distributed with the training set after training on millions of labeled images. However, when they are deployed in a target domain that has a significantly different domain distribution from the source domain, their performance usually drops sharply. This mismatched domain distribution is called domain shift. Domain shift is caused by various perspectives, various lighting conditions, and different sensors, but can be minimized by using transfer learning methods. One intuitive method is to manually label the target domain images and fine-tune the already trained model. However, labeling a large number of labels on the target domain is time-consuming and expensive.

[0003] To solve this problem, unsupervised domain adaptation (UDA) methods propose to transfer label knowledge from the labeled source domain to the unlabeled target domain. The classic scheme of unsupervised domain adaptation is to use domain hard labels (i.e., mark source domain images as 1 and target domain images as 0), and then apply a gradient reverse layer (GRL) to minimize the difference in marginal distribution between the source domain and the target domain.

[0004] In real-world scenarios, some images in the target domain may not belong to any class in the source domain, i.e., unknown classes. Recently proposed open-set domain adaptation (OSDA) proposes that, on the basis of unsupervised domain adaptation, the model also needs to distinguish unknown class images that do not belong to any class in the source domain. In this case, if the entire target domain is forcibly aligned with the source domain, negative transfer will inevitably occur. Domain hard labels will weaken the performance of the model, and directly using domain hard labels for open-set domain adaptation will align the entire target domain with the source domain, so unknown class images will be incorrectly predicted by the classification head as common classes shared by the source domain and the target domain.

[0005] Therefore, it is necessary to design a new learning paradigm for open-set image classification field adaptation, so that the model can accurately distinguish common classes and common classes and perform domain adaptation, which is an urgent problem for those skilled in the art to overcome. SUMMARY

[0006] The present application is just for the deficiencies of the application of the existing unsupervised domain adaptation in the actual scene, and provides an open set image classification domain adaptation method based on self-paced learning. First, the original image is preprocessed to obtain an image set, and then a feature extraction module and a double multi-class classifier module are constructed and trained to align the shared class features of the source domain image and the target domain image and separate the private class features of the target domain image. Then, a multi-criteria cross-domain hybrid module is constructed and trained, and the shared class features are self-paced learned by generating cross-domain hybrid images from the source domain image and the target domain image. Finally, the classification result of the target domain image is output. The present application overcomes the shortcomings of the existing unsupervised domain adaptation model that cannot well distinguish the common class and the unknown class. Compared with the existing open set image classification domain adaptation method, the present application covers smooth and non-smooth class distribution, and does not need to empirically adjust the threshold for distinguishing the common class image and the private class image in the inference stage, so that the model has good robustness under different hyperparameters and experimental settings.

[0007] To achieve the above object, the technical scheme adopted by the present application is as follows: an open set image classification domain adaptation method based on self-paced learning, comprising the following steps:

[0008] S1, image set acquisition: the original image is preprocessed to obtain an image set, wherein the image set includes a source domain image set with artificial annotation labels and a target domain image set without artificial annotation labels;

[0009] S2, construction and training of a feature extraction module and a double multi-class classifier module: the feature extraction module is composed of a feature extractor F, and the double multi-class classifier module is composed of an adversarial classifier G C and an auxiliary classifier G aux The feature extractor F, the adversarial classifier G C and the auxiliary classifier G aux are used to align the shared class features of the source domain image and the target domain image and separate the private class features of the target domain image.

[0010] S3, construction and training of a multi-criteria cross-domain hybrid module: the multi-criteria cross-domain hybrid module is composed of m classifiers , and the shared class features are self-paced learned by generating cross-domain hybrid images from the source domain image and the target domain image obtained in step S1.

[0011] S4, result output: the target domain image is input, and the classification result of the target domain image is output by the feature extraction module, the adversarial classifier G C and the multi-criteria cross-domain hybrid module trained in steps S2 and S3.

[0012] As an improvement of the present application, in step S1, the source domain image set with artificial annotation labels whose class set is C S a target domain image set without artificial labeled tags whose class set is C T wherein is a source domain image, is a source domain image a corresponding label, i, j is a positive integer greater than or equal to 1, indicating the serial number of the image, N S is the total number of source domain images, N T is the total number of target domain images, C S is C T a proper subset; openness wherein |·| represents the power of a set.

[0013] As an improvement of the present application, the method of aligning the shared class features of the source domain images and the target domain images and separating the target domain private class features in the step S2 is specifically as follows:

[0014] the adversarial classifier G C and the auxiliary classifier G aux jointly calculate the probability that a given image x belongs to a common class

[0015]

[0016] wherein, denotes the summation of the first |C C | elements of the probability vector G S (x), denotes the summation of the |C aux | elements of the probability vector G S (x);

[0017] the shared class features are aligned and the target domain private class features are separated by using P common (x) and G C , and the loss function is as follows:

[0018]

[0019] wherein, denotes the (|C C |+1)th element of the probability vector G t output by the adversarial classifier G C (x) for the target domain image x S .

[0020] As another improvement of the present application, in the step S2, the optimization objectives of the feature extractor F, the adversarial classifier G C and the auxiliary classifier G aux are respectively as follows:

[0021]

[0022]

[0023]

[0024] wherein, denotes the adversarial classifier G C the classification loss on source domain images; denotes the auxiliary classifier G aux the classification loss on source domain images; denotes the auxiliary classifier G aux the domain classification loss on source domain images and target domain images.

[0025] As another improvement of the present application, the adversarial classifier G C the classification loss on source domain images wherein y s is the corresponding true label of the image x s , is the standard cross-entropy loss function;

[0026] the auxiliary classifier G aux the classification loss on source domain images wherein y' s is the true label in one-hot form, is the standard binary cross-entropy loss function;

[0027] the auxiliary classifier G aux the domain classification loss on source domain images and target domain images wherein ||·|| denotes the kernel norm, G * (x) is the output of G aux , |C aux | is the number of classes, and S |C s | is the dimension of the probability vector.

[0028] As yet another improvement of the present application, the loss function of the step S3 is:

[0029]

[0030] wherein λ2 is a hyperparameter controlling the mixing proportion of cross-domain images, y s is the corresponding true label of the source domain image x s , is the pseudo label about the target domain image x C predicted by the adversarial classifier G t , h is a self-computed threshold, and ω tis a multi-criteria coefficient.

[0031] As a further improvement of the present application, the optimization target of the step S3 multi-criteria cross-domain hybrid module is:

[0032]

[0033] As a further improvement of the present application, in the step S4, the trained feature extraction module, the adversarial classifier G C and the multi-criteria cross-domain hybrid module output the classification results of the target domain image

[0034]

[0035] wherein |C S |+1 represents the unknown class number, represents the pseudo label output by the adversarial classifier G C .

[0036] Compared with the prior art, the present application has the technical advantages and effects that: the present application provides an open set image classification field adaptation method based on self-paced learning, a double classifier module is designed to accurately align the shared class features of the source domain image and the target domain image and separate the target domain private class features, avoiding the negative transfer phenomenon that the classifier incorrectly identifies unknown class images as common class images due to aligning the entire target domain with the source domain in the unsupervised image field adaptation scheme; at the same time, the present method contains a self-paced learning method based on multi-criteria cross-domain image mixing, which can learn domain-invariant features in a more continuous feature space and make the model more accurately distinguish common class and unknown class samples. Compared with the existing open set image classification field adaptation method, the present method covers smooth and non-smooth class distributions, and does not need to empirically tune the threshold for distinguishing common class images and private class images in the inference stage, so that the model has good robustness under different hyperparameters and experimental settings. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a schematic diagram of the open set image classification field adaptation scenario of the present application;

[0038] Figure 2 is a step flowchart of the open set image classification field adaptation method based on self-paced learning of the present application;

[0039] Figure 3 is a model structure framework diagram of the training stage of the method step S2 of the present application;

[0040] Figure 4 ​​A cross-domain image mixing schematic diagram in the training stage of the multi-criteria cross-domain hybrid module in the method of the present application;

[0041] Figure 5 A sample graph showing the change of the model self-computed threshold h with the openness O of the data set and the number of iterations in the method of the present application;

[0042] Figure 6 A change schematic diagram of the cross-domain mixed image in the method of the present application;

[0043] Figure 7 A model framework diagram in the inference stage of the method of the present application. DETAILED DESCRIPTION

[0044] The present application will be further illustrated below in conjunction with the accompanying drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0045] Embodiment 1

[0046] For the convenience of understanding, the following explains the terms involved in the present embodiment.

[0047] (1) Gradient Reversal Layer: The gradient reversal layer is a technique commonly used in domain adaptation, which is used to reduce the difference between the source domain and the target domain. The main role of the gradient reversal layer is to suppress the difference between the source domain and the target domain, so as to realize the domain adaptation of the model. The gradient reversal layer is usually a layer of neural network, which passes the input feature vector to the next layer without change, but multiplies the weight of the gradient back propagation by a negative number, so as to offset the influence of the original gradient. In this way, the gradient reversal layer can model the difference between the source domain and the target domain in the training process, and reduce the influence of these differences.

[0048] (2) Self-paced learning: Self-paced learning is a training strategy in machine learning that aims to help the model learn and generalize better by gradually increasing the difficulty and complexity of training samples. In self-paced learning, training data is learned in a certain order, which is automatically adjusted by the model based on its learning progress and capabilities. Initially, the model only learns easily identifiable data, such as data with low noise and obvious categories. As the model's learning progresses to a certain extent, the complexity and difficulty of the training data are gradually increased. Self-paced learning can help machine learning models better adapt to complex training data and improve their generalization ability. Compared to traditional training methods, self-paced learning can reduce the model's sensitivity to noise and outliers, thus better handling the uncertainty and diversity in the real world. Self-paced learning can be applied to various machine learning tasks, including classification, regression, and clustering.

[0049] (3) Inference phase: The inference phase refers to the process after the machine learning model has been trained, where new input data is fed into the model to obtain prediction results. During the inference phase, the model no longer updates its parameters, but instead uses the previously trained parameters and structure to predict or classify new data.

[0050] Example 1

[0051] like Figure 1 As shown, in the adaptive scenario of open-set image classification, source domain images have manually labeled ground truth labels, while target domain images do not have manually labeled ground truth labels; the set of image categories in the source domain is C. S The set of image categories in the target domain is C. T C S and C T The relationship between them is C S It is C T A proper subset of the class set C = C S ∩C T The categories in a set of categories are called common classes. The categories in the target domain are called unknown classes. The common classes are those shared by the source domain and the target domain. The unknown classes are those possessed by the target domain but not by the source domain. In the adaptive task of open set image classification, it is necessary to align the common class features of the target domain and the source domain and separate the unknown class features of the target domain from those of the source domain.

[0052] An open-set image classification domain adaptation method based on self-stepping learning, such as Figure 2 As shown, it includes the following steps:

[0053] Step S1, image set acquisition: acquire an image set, which contains source domain images with artificial annotation labels and a class set C S and target domain images without artificial annotation labels and a class set C T is a source domain image, is a source domain image corresponding label, i, j are positive integers greater than or equal to 1, used to indicate the serial number of the image, N S is the total number of source domain images, N T is the total number of target domain images, C S is a proper subset of C T ; openness

[0054] The data distribution of the source domain image x s is p, and the data distribution of the target domain image x t is q; wherein p and q have certain distribution differences, so that the source domain image x s and the target domain image x t cannot be regarded as independent and identically distributed image sets; for the classes in the class set C = C S ∩ C T , the classes are called common classes, and for the classes in the class set , the classes are called unknown classes, the common classes are classes common to the source domain and the target domain, and the unknown classes are classes possessed by the target domain and not possessed by the source domain.

[0055] The method for acquiring the image set in the training stage comprises: providing original images; and performing image preprocessing on the original images. Preferably, the image preprocessing method in the training stage comprises: resizing the original images to images with a size of 256x256 pixels; performing horizontal rotation on the resized images with a probability of p = 0.5; randomly cropping a region with a size of 224x224 pixels from the images with a size of 256x256 pixels; and performing normalization on the cropped region images.

[0056] Step S2, constructing and training a feature extraction module and a dual multi-class classifier module: the feature extraction module is composed of a feature extractor F, and the dual multi-class classifier module is composed of an adversarial classifier G C and an auxiliary classifier G aux , each module is as shown in Figure 3 the feature extractor F, the adversarial classifier G C and the auxiliary classifier G aux ​​​​aligning shared class features of source domain images and target domain images and separating target domain private class features;

[0057] The shared class features are features of classes that are common to the source domain images and the target domain images, and the target domain private class features are features of classes that the target domain images have and the source domain images do not have. C A threshold value h is calculated using the target domain images.

[0058] In this embodiment, the feature extractor F is a ResNet-50 convolutional neural network plus a 2048x1000-dimensional fully connected layer, the input of the feature extractor F is an image x of 224x224 pixels, and the output is an image feature z.

[0059] The input of the adversarial classifier G C is the image feature z, and the output is a logit vector of (|C S |+1) dimensions After normalization by a softmax function, a probability vector can be obtained:

[0060]

[0061] |C S | is the number of classes in the source domain, represents the sum of the (|C S |+1) elements of the logit vector;G C The (|C S |+1)th element of the function f(x) indicates the probability that the image x belongs to the unknown class, and the remaining |C S | elements indicate the probability that the image x belongs to each corresponding shared class.

[0062] The input of the auxiliary classifier G aux is the image feature z, and the output is a logit vector of |C S | dimensions After normalization by a leakysoftmax function, a probability vector can be obtained:

[0063]

[0064] |C S | is the number of classes in the source domain;G aux The |C S | elements of the function f(x) indicate the probability that the image x belongs to each corresponding shared class.

[0065] The classification loss of the adversarial classifier G C on the source domain images is:​

[0066]

[0067] y s is the image x s corresponding real label, is the standard cross-entropy loss function.

[0068] the auxiliary classifier G aux the classification loss on the source domain images is:

[0069]

[0070] y′ s is the real label in one-hot form, is the standard binary cross-entropy loss function.

[0071] the auxiliary classifier G aux the domain classification loss on the source domain images and the target domain images is:

[0072]

[0073] ||·|| * represents the nuclear norm, G aux (x) is the G aux output |C S | dimensional probability vector.

[0074] the adversarial classifier G C and the auxiliary classifier G aux jointly compute the probability that the given image x belongs to the common class:

[0075] P common (x) = P1(x) x P2(x),

[0076]

[0077] represents the summation of the first |C C | elements of the probability vector G S (x), represents the summation of the |C aux | elements of the probability vector G S (x).

[0078] align the shared class features and separate the target domain private class features using P common (x) and G C , and the loss function is:

[0079]

[0080] represent G C for the target domain image x t output probability vector G C (|C S |+1)th element of G

[0081] the feature extractor F, the adversarial classifier G C and the auxiliary classifier G aux The optimization objective includes:

[0082] The optimization objective is constructed using equations (3) and (7) optimizing the adversarial classifier G C ;

[0083] The optimization objective is constructed using equations (3) and (7) optimizing the feature extractor F;

[0084] The optimization objective is constructed using equations (4) and (5) optimizing the auxiliary classifier G aux .

[0085] The threshold h is calculated using the target domain image as the input of the adversarial classifier G C , and the calculation formula is:

[0086]

[0087] i, j are the serial numbers of randomly extracted source domain images, λ1∈[0.5, 1.0] is a hyperparameter artificially set, and ⊙ represents dot multiplication. representing the |C C |+1th element of the probability vector G t (x S ) is discarded, and only the first |C S | elements are used for vector addition and dot multiplication.

[0088] Step S3, constructing and training a multi-criteria cross-domain hybrid module: the multi-criteria cross-domain hybrid module is composed of m classifiers , and the source domain images and target domain images obtained in step S1 are used to generate cross-domain hybrid images and learn shared class features;

[0089] The input of the m classifiers is the image feature z, and the output is a logit vector of |C S | dimensions After normalization by the softmax function, the probability vector

[0090] ​

[0091] |C S | is the number of classes in the source domain; |C S | elements indicate the probability that the image x belongs to the corresponding shared class.

[0092] For the target domain image x t , the probability vector output by the m classifiers in the multi-criteria cross-domain hybrid module is used to calculate the entropy ω ent , consistency ω cons and confidence ω conf , including:

[0093]

[0094]

[0095]

[0096] The cth element of represents the maximum element value of the probability vector .

[0097] In combination with equations (10), (11) and (12), the multi-criteria coefficient t of the target domain image x

[0098]

[0099] As shown in equation (13), the classifiers in the multi-criteria cross-domain hybrid module Figure 4 generate cross-domain hybrid images using the source domain images and the target domain images and self-learn the shared class features, and the loss function is:

[0100]

[0101] λ2 is a hyperparameter that controls the proportion of cross-domain image mixing, y s is the real label corresponding to the source domain image x s , is the pseudo label about the target domain image x C predicted by the adversarial classifier G t , h is the threshold value calculated by equation (8), ω t is the multi-criteria coefficient calculated by equation (13). Figure 5 ​​The model self-computes the threshold h with the changes of the openness 0 of the data set and the number of iterations; wherein the numbers in the brackets after the data sets Office-31 and OfficeHome represent the number of classes of the shared classes of the source domain and the target domain and the number of classes of the unknown classes of the source domain. Taking the Office-31 data set as an example, the Office-31 data set has a total of 31 classes, and the legend "Office-31 (10 / 21)…" represents that 10 of them are divided into shared classes, and the remaining 21 are divided into unknown classes; when the number of classes of the shared classes increases and the number of classes of the unknown classes decreases, the openness O of the data set will decrease. As shown in Figure 5 , the self-computed threshold h is higher in the initial stage of training, thereby reducing the noise of the pseudo-labels when the multi-criteria cross-domain hybrid module self-studies; at the same time, as the openness O of the data set decreases, the target domain image will contain more shared class images, so under the same data set open set domain migration task and the same number of iterations, the threshold h self-computed by the model according to formula (8) is lower, so that the model can generate more rich cross-domain hybrid images. Figure 6 The changes of the cross-domain hybrid images of the model at different stages of the training stage are shown in Figure 6 , in the early stage of the training stage, the model only uses simpler target domain images and pseudo-labels to generate fewer cross-domain hybrid images, and as the training progresses, the threshold is gradually adjusted to a suitable value, so that the model uses more complex target domain images and pseudo-labels to generate more cross-domain hybrid images in the later stage of the training stage. Therefore, the model can gradually learn the domain-invariant features from the simpler images to the more complex images in the target domain. At this time, due to the decrease of the threshold, the number of target domain images with pseudo-labels is increased, and the noise brought by the pseudo-labels is increased, but since the learning rate of the model is already relatively low, the negative impact on the performance of the model is small.

[0102] S4, result output: taking the target domain image as input, using the feature extraction module, the adversarial classifier G C and the multi-criteria cross-domain hybrid module trained in steps S2 and S3 to output the classification result of the target domain image.

[0103] Figure 7 The inference stage of an open set image classification domain adaptation method based on self-study is shown, which uses the trained feature extraction module, adversarial classifier G C and multi-criteria cross-domain hybrid module to output the classification result of the target domain image

[0104]

[0105] |C S ​|+1 represents an unknown class number, representing the adversarial classifier G C For Output pseudo-labels.

[0106] Test example

[0107] Office-31 is a benchmark data set for domain adaptation, containing 31 object categories in 3 domains of Amazon (A), DSLR (D) and Webcam (W). We select the first 10 categories in alphabetical order as the shared classes of the source domain and the target domain, and the remaining 21 categories as unknown classes of the target domain.

[0108] Table 10 OS (%) and HOS (%) on the Office-31 data set

[0109]

[0110]

[0111] OS is used to evaluate the average accuracy of the model in |C S |+1 categories, including the accuracy of the model in |C S | shared classes and the accuracy of 1 total unknown class. HOS is the harmonic mean of the average accuracy OS* on |C S | shared classes and the accuracy Unk on 1 total unknown class. The formal definition is as follows:

[0112]

[0113] The Ours method shown in Table 1 is the method of the present application, and it can be seen from the table that the method of the present application is compared with the best 9 methods applicable to the field of open set image classification domain adaptation, and the best performance is achieved on the Office-31 data set, the OS score is improved by 1.2 percentage points compared with the best existing method OSBP, and the HOS score is improved by 2.09 percentage points compared with the best existing method UADAL method, proving the effectiveness of the present application.

[0114] In summary, the present application provides an implementation method of an open set image classification domain adaptation model based on self-paced learning, the core of which is a dual multi-class classifier module and a multi-criteria cross-domain hybrid module. The method of the present application is convenient, fast, effective, can improve the migration efficiency and the accuracy of model prediction, and has high popularization value.

[0115] It should be noted that the above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. For ordinary skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can be made, which fall within the protection scope of the claims of the present application.

Claims

1. A domain-adaptive open-set image classification method based on self-stepping learning, characterized in that, Includes the following steps: S1, Image set acquisition: Preprocess the original image to obtain an image set, which includes a source domain image set with manually labeled images and a target domain image set without manually labeled images. S2, Constructing and training the feature extraction module and the dual multi-class classifier module: The feature extraction module consists of a feature extractor... The dual multi-class classifier module consists of an adversarial classifier. and auxiliary classifier Composition, Feature Extractor Adversarial classifier and auxiliary classifier This method is used to align shared category features between source and target domain images and separate private category features of the target domain; specifically, the method for aligning shared category features between source and target domain images and separating private category features of the target domain is as follows: Adversarial classifier and auxiliary classifier Joint computation of a given image Probability of belonging to a common class ； in, This indicates that for a probability vector The former Summing the elements of the terms This indicates that for a probability vector of Summing the elements of the terms; use and Align shared class features and separate target domain private class features; the loss function is: ； in, express For the target domain image Output probability vector The Item element; Feature extractor Adversarial classifier and auxiliary classifier The optimization objectives are as follows: ； ； ； in, Adversarial classifier Classification loss on the source domain image; Auxiliary classifier Classification loss on the source domain image; Represents the auxiliary classifier Domain classification loss on source and target domain images; S3, Constructing and training a multi-criteria cross-domain hybrid module: The multi-criteria cross-domain hybrid module is composed of... Classifier The process involves generating a cross-domain hybrid image using the source domain image and target domain image obtained in step S1, and then automatically learning shared category features. S4, Output: Taking the target domain image as input, the feature extraction module and adversarial classifier trained in steps S2 and S3 are used. The multi-criteria cross-domain hybrid module outputs the classification results of the target domain image.

2. The open-set image classification domain adaptation method based on self-synchronization learning as described in claim 1, characterized in that: In step S1, there is a set of source domain images with manually labeled tags. The category set is Collection of target domain images without manual annotation Its category set is ,in It is a source domain image. It is a source domain image The corresponding tags It is a positive integer greater than or equal to 1, indicating the image's sequence number. It is the total number of images in the source domain. It is the total number of images in the target domain. yes A proper subset of; Openness ,in It represents the cardinality of a set.

3. The open-set image classification domain adaptation method based on self-synchronization learning as described in claim 2, characterized in that: Adversarial classifier Classification loss on source domain image ,in, It is an image The corresponding real tags, It is the standard cross-entropy loss function; Auxiliary classifier Classification loss on source domain image ,in, It is a true label in one-hot format. It is the standard binary cross-entropy loss function; Auxiliary classifier Domain classification loss on source and target domain images ,in, Represents the nuclear norm. yes Output A probability vector of dimension.

4. The open-set image classification domain adaptation method based on self-synchronization learning as described in claim 3, characterized in that: The loss function in step S3 is: ； It is a hyperparameter that controls the blending ratio of cross-domain images. It is a source domain image The corresponding real tags, It is an adversarial classifier Predicted image of the target domain pseudo-tags It is a self-calculated threshold. It is a multi-criteria coefficient.

5. The open-set image classification domain adaptation method based on self-synchronization learning as described in claim 4, characterized in that: The optimization objective of the multi-criteria cross-domain hybrid module in step S3 is:

6. The open-set image classification domain adaptation method based on self-synchronization learning as described in claim 5, characterized in that: In step S4, the trained feature extraction module and adversarial classifier are utilized. The multi-criteria cross-domain hybrid module outputs the target domain image. Classification results : ； in, Indicates the unknown class index. This represents the adversarial classifier. for Output pseudo-tags.

Images