A foreground object transfer method for unsupervised domain adaptation

Abstract

The application discloses a foreground target migration method for unsupervised domain adaptation, and the method comprises the following steps: obtaining labeled source domain samples and unlabeled target domain samples as training samples, inputting the source domain samples and the target domain samples into a deep neural network for training to obtain a classification model, inputting data to be classified in the target domain into the classification model, and obtaining a classification result. The application further provides a foreground target migration device for unsupervised domain adaptation. The foreground target migration method and device for unsupervised domain adaptation provided by the application can improve classification precision and classification efficiency.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a foreground target transfer method for unsupervised domain adaptation. Background Technology

[0002] Unsupervised domain adaptation, based on deep neural networks, involves training a classification model from a labeled source domain to adapt to an unlabeled target domain. However, due to the different data distributions in the source and target domains, the classification model performs poorly in the target domain.

[0003] In existing technologies, the performance of classification models is improved by aligning global features of the source and target domains and learning domain-invariant features. However, this approach ignores the differences between foreground and background features in the samples and does not consider the structural information in the foreground objects of the samples, i.e., it ignores the fine class-specific structure of the samples, resulting in noisy predictions near the boundaries of the classification model.

[0004] Therefore, it is necessary to provide an unsupervised adaptive method to improve classification accuracy and efficiency. Summary of the Invention

[0005] The purpose of this application is to improve classification accuracy and efficiency.

[0006] To achieve the above objectives, embodiments of the present invention provide a method and apparatus for foreground target transfer in unsupervised domain adaptation. The technical solution is as follows:

[0007] Firstly, a foreground target transfer method for unsupervised domain adaptation, the method comprising:

[0008] Obtain labeled source domain samples and unlabeled target domain samples as training samples;

[0009] The source domain samples and the target domain samples are input into a deep neural network for training to obtain a classification model;

[0010] The data to be classified in the target domain is input into the classification model to obtain the classification result.

[0011] Furthermore, the method also includes:

[0012] Extract foreground features from the source domain samples and extract foreground features from the target domain samples;

[0013] The foreground features in the source domain samples and the foreground features in the target domain samples are compared and learned.

[0014] Further, extracting foreground features from the source domain samples and extracting foreground features from the target domain samples includes:

[0015] Based on prior knowledge, foreground features and background features in the source domain sample or the target domain sample are distinguished.

[0016] Further, after extracting the foreground features from the source domain samples, the process includes:

[0017] Set the weights of the foreground features in the source domain samples.

[0018] Furthermore, based on the foreground features in the source domain samples and the foreground features in the target domain samples, comparative learning is performed, including:

[0019] A class-level alignment method is used to compare the foreground features in the source domain sample with the foreground features in the target domain sample.

[0020] Secondly, a foreground target migration apparatus for unsupervised domain adaptation, the apparatus comprising:

[0021] The extraction module is used to obtain labeled source domain samples and unlabeled target domain samples as training samples;

[0022] The training module is used to input the source domain samples and the target domain samples into a deep neural network for training to obtain a classification model;

[0023] The classification module is used to input the data to be classified in the target domain into the classification model to obtain the classification result.

[0024] Furthermore, the extraction module is also used for:

[0025] Extract foreground features from the source domain samples and extract foreground features from the target domain samples;

[0026] The foreground features in the source domain samples and the foreground features in the target domain samples are compared and learned.

[0027] Furthermore, the extraction module is also used for:

[0028] Based on prior knowledge, foreground features and background features in the source domain sample or the target domain sample are distinguished.

[0029] Furthermore, the extraction module is also used for:

[0030] Set the weights of the foreground features in the source domain samples.

[0031] Furthermore, the extraction module is also used for:

[0032] A class-level alignment method is used to compare the foreground features in the source domain sample with the foreground features in the target domain sample.

[0033] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows: In the process of unsupervised domain adaptation, prior knowledge is used to distinguish and compare the foreground features and background features in the source domain samples and the target domain samples, taking into account the structural information in the foreground objects, thereby avoiding noise prediction near the boundary of the classification model and improving the classification accuracy and classification efficiency of the classification model. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0035] Figure 1 This is a schematic diagram of a foreground target migration method for unsupervised domain adaptation in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of a foreground target migration device for unsupervised domain adaptation in an embodiment of the present invention. Detailed Implementation

[0037] 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.

[0038] It should be noted that due to the development of deep convolutional networks, unsupervised domain adaptation methods such as adversarial learning-based unsupervised domain adaptation and metric-based unsupervised domain adaptation have emerged. However, adversarial learning-based unsupervised domain adaptation methods do not utilize source domain label information, making them class-agnostic methods. This means they tend to overlook the more refined class-specific structure of samples, leading to noisy predictions near the classification model boundaries.

[0039] In view of this, please refer to Figure 1This invention discloses a foreground target transfer method for unsupervised domain adaptation, wherein the method possesses only unlabeled target domain samples and a pre-trained model trained on the source domain. Then, all unlabeled target domain samples are input into the source domain model, and an accuracy threshold is set. Based on this threshold, the target domain samples are divided into a high-accuracy sample set T1 and a low-accuracy sample set T2. Since T1 has higher accuracy, the predicted values ​​of T1 are used as the true labels. That is, the method possesses labeled T1, unlabeled T2, and the pre-trained model. Specifically, the method may include the following steps:

[0040] S101: Obtain labeled source domain samples and unlabeled target domain samples as training samples.

[0041] S102: Input the source domain samples and the target domain samples into a deep neural network for training to obtain a classification model.

[0042] S103: Input the data to be classified in the target domain into the classification model to obtain the classification result.

[0043] In practice, labeled source domain samples and unlabeled target domain samples are obtained separately, i.e., labeled source domain samples and unlabeled target domain samples are obtained, which are used as training samples and input into the neural network for training to obtain a classification model.

[0044] In one optional implementation, the training samples actually include foreground features from the source domain samples and foreground features from the target domain samples. Specifically, structural information in the source and target domain samples is obtained by extracting the positional information of shallow features from the samples. Subsequent use of prior information is simply to further extract features related to unsupervised domain adaptation from the structural information. Furthermore, since the label information in the target domain samples is originally missing, pseudo-labels can be assigned to the target domain samples using a clustering algorithm.

[0045] In an alternative implementation, to avoid the impact of noisy background information on the unsupervised domain adaptation performance, the weights of the foreground features in the source domain samples are set, i.e., the weights of the foreground features in the source domain samples are increased to align with the foreground features in the target domain samples. This implementation can also avoid situations where depth features are both positive features that are conducive to alignment and negative features that are not conducive to alignment when comparing the foreground features in the source domain samples with the foreground features in the target domain samples, i.e., avoid situations that lead to negative alignment.

[0046] In one optional implementation, class-level alignment is used to compare foreground features in the source domain samples with foreground features in the target domain samples. To compare the structural information of the foreground features in the source and target domains, it is necessary to preserve the relative positional relationships of different parts of the sample image, i.e., to preserve shallow features in the sample image. Specifically, high-resolution feature maps in the sample image are sampled, then merged with low-resolution feature maps, and all are added to the depth feature map, thereby preserving the structural information in the sample foreground features to a greater extent. Alternatively, pooling and simple joins can also be used to preserve the positional relationships of different parts of the sample image.

[0047] In this embodiment, before comparing the foreground features in the source domain samples with the foreground features in the target domain samples, a contrastive loss function is used to minimize the foreground features between similar samples in the source domain samples and the target domain samples, and to maximize the difference between foreground features of different categories in the source domain samples and the target domain samples. This achieves positive alignment of similar foreground features in the source domain samples and the target domain samples, while increasing the geometric distance between foreground features of different categories in the source domain samples and the target domain samples. This prevents the influence of negative features during the training of the classification model, thereby improving the classification accuracy of the classification model.

[0048] Based on the same concept, embodiments of the present invention also provide, for example... Figure 2 The illustrated device is a foreground target transfer apparatus for unsupervised domain adaptation, the apparatus comprising:

[0049] The extraction module is used to obtain labeled source domain samples and unlabeled target domain samples as training samples;

[0050] The training module is used to input the source domain samples and the target domain samples into a deep neural network for training to obtain a classification model;

[0051] The classification module is used to input the data to be classified in the target domain into the classification model to obtain the classification result.

[0052] Furthermore, the extraction module is also used for:

[0053] Extract foreground features from the source domain samples and extract foreground features from the target domain samples;

[0054] The foreground features in the source domain samples and the foreground features in the target domain samples are compared and learned.

[0055] Furthermore, the extraction module is also used for:

[0056] Based on prior knowledge, foreground features and background features in the source domain sample or the target domain sample are distinguished.

[0057] Furthermore, the extraction module is also used for:

[0058] Set the weights of the foreground features in the source domain samples.

[0059] Furthermore, the extraction module is also used for:

[0060] A class-level alignment method is used to compare the foreground features in the source domain sample with the foreground features in the target domain sample.

[0061] As can be seen from the above, the unsupervised domain adaptation method and apparatus provided in the embodiments of the present invention. During the unsupervised domain adaptation process, prior knowledge is used to distinguish and compare foreground and background features in source and target domain samples, taking into account the structural information in the foreground object. This avoids noise prediction near the boundary of the classification model, thereby improving the classification accuracy and efficiency of the classification model.

[0062] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0063] Through the description of the above-described embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus necessary general-purpose hardware platforms, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A foreground target transfer method for unsupervised domain adaptation, characterized in that, The method includes: Obtain labeled source domain samples and unlabeled target domain samples as training samples; Based on prior knowledge, foreground features and background features in the source domain samples are distinguished, and foreground features in the source domain samples are extracted, and the weight of the foreground features in the source domain samples is increased. Based on prior knowledge, foreground features and background features in the target domain sample are distinguished, and foreground features in the target domain sample are extracted; The foreground features in the source domain samples and the foreground features in the target domain samples are input into a deep neural network, which then learns and trains based on the comparison between the foreground features in the source domain samples and the foreground features in the target domain samples to obtain a classification model. The data to be classified in the target domain is input into the classification model to obtain the classification result.

2. The method according to claim 1, characterized in that, Based on the foreground features in the source domain samples and the foreground features in the target domain samples, comparative learning is performed, including: A class-level alignment method is used to compare the foreground features in the source domain sample with the foreground features in the target domain sample.

3. A foreground target transfer device for unsupervised domain adaptation, characterized in that, The device includes: The extraction module is used to acquire labeled source domain samples and unlabeled target domain samples as training samples; it is also used to distinguish foreground features and background features in the source domain samples based on prior knowledge, and extract the foreground features in the source domain samples, increasing the weight of the foreground features in the source domain samples; and it is used to distinguish foreground features and background features in the target domain samples based on prior knowledge, and extract the foreground features in the target domain samples. The training module is used to input the foreground features in the source domain samples and the foreground features in the target domain samples into the deep neural network, so that the deep neural network learns and trains based on the comparison between the foreground features in the source domain samples and the foreground features in the target domain samples to obtain a classification model; The classification module is used to input the data to be classified in the target domain into the classification model to obtain the classification result.

4. The apparatus according to claim 3, characterized in that, The extraction module is also used for: A class-level alignment method is used to compare the foreground features in the source domain sample with the foreground features in the target domain sample.

Images