A self-training method based on hybrid domain and co-trained cross-domain detection models

By using a cross-domain detection model that combines hybrid domains and co-training, the problems of domain discrepancies and low quality of pseudo-labels are solved, improving image detection accuracy and network performance, and generating higher quality pseudo-labels.

CN115661565BActive Publication Date: 2025-10-31HEFEI UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202211120391.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-15
Publication Date
2025-10-31
Estimated Expiration
2042-09-15

AI Technical Summary

Technical Problem

In existing technologies, domain differences and low quality of pseudo-labels lead to reduced target detection accuracy.

Method used

A cross-domain detection model based on hybrid domain and co-training is adopted. A hybrid domain dataset is generated through CycleGAN network, and feature extraction and pseudo-label labeling are performed using the co-training model. High-quality pseudo-labels are selected by combining position threshold and category threshold, and model parameters are optimized.

Benefits of technology

It improved image detection accuracy, increased the number of samples, enhanced network performance, generated higher quality pseudo-labels, and optimized model performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115661565B_ABST
    Figure CN115661565B_ABST
Patent Text Reader

Abstract

This invention belongs to the field of deep learning and object detection, specifically relating to a self-training method for a cross-domain detection model based on hybrid domain and co-training. The method includes: collecting a source domain dataset and a target domain dataset; inputting the source domain dataset and the target domain dataset into a CycleGAN network to obtain a hybrid domain dataset; inputting the hybrid domain dataset into a co-training model for training to obtain an initial training model; inputting the target domain dataset into the initial training model to obtain the location and category information of the target domain and label it as a pseudo-label of the target domain, while also calculating model optimization parameters; and inputting the pseudo-labeled target domain dataset, model optimization parameters, and hybrid domain dataset into the initial training model for training to obtain a cross-domain detection model based on hybrid domain and co-training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of deep learning and object detection, specifically relating to a self-training method for a cross-domain detection model based on hybrid domain and collaborative training. Background Technology

[0002] Object detection is a crucial branch of computer vision with wide-ranging applications. The advent of deep learning network models has further propelled the development of this field. Deep learning models adaptively learn from images, representing an end-to-end detection method. With the advent of the big data era, the continuous enrichment and improvement of datasets used to train deep learning network models have also driven the development of deep learning-based computer vision. The number of labeled samples has a crucial impact on deep learning, directly affecting network performance; however, collecting source domain images with labeled information is expensive and sometimes impossible.

[0003] In deep learning, a domain refers to the scene of an image. For example, daytime and darkness are different domains, as are sunny days and foggy days. Objects exhibit different characteristics in different domains. However, in the task of domain adaptation for object detection, due to the differences between domains, models trained in the source domain often experience a significant drop in accuracy when tested in the target domain.

[0004] In mainstream object detection domain adaptation tasks, single-step and multi-step domain adaptation methods are employed. Single-step methods typically involve directly training the source and target domains. Multi-step methods, on the other hand, generally first use domain transfer to generate intermediate and target domains, essentially providing pseudo-labels for the unlabeled target domain during training. This multi-step domain adaptation method using pseudo-labels can produce considerable results, but it focuses only on the target and generated intermediate domains during training, neglecting the source domain. Furthermore, the quality of pseudo-labels generated by a single network is limited, and the quality of these pseudo-labels has a crucial impact on model performance, thus limiting its effectiveness. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a self-training method based on a hybrid domain and co-trained cross-domain detection model, which can solve the technical problem of reduced image detection accuracy caused by domain differences or low quality of pseudo-labels.

[0006] To achieve the above and other related objectives, this invention provides a self-training method for a cross-domain detection model based on hybrid domain and co-training, comprising: collecting a source domain dataset and a target domain dataset; inputting the source domain dataset and the target domain dataset into a CycleGAN network to obtain a hybrid domain dataset; inputting the hybrid domain dataset into a co-training model for training to obtain an initial training model; inputting the target domain dataset into the initial training model, extracting feature parameters from the target domain in the target domain dataset through a feature extraction network F, and then inputting the feature parameters into branch modules F1 and F2 to obtain the position and category information of the target domain and label it as a pseudo-label of the target domain, while also calculating model optimization parameters; inputting the target domain dataset labeled with pseudo-labels, the model optimization parameters, and the hybrid domain dataset into the initial training model to obtain a cross-domain detection model based on hybrid domain and co-training.

[0007] According to a specific embodiment of the present invention, the acquisition of the source domain dataset and the target domain dataset further includes: performing size normalization processing on the source domain dataset and the target domain dataset.

[0008] According to a specific embodiment of the present invention, the step of inputting the source domain dataset and the target domain dataset into the CycleGAN network to obtain a hybrid domain dataset includes: step S21, inputting the source domain dataset and the target domain dataset into the CycleGAN network; step S22, calculating a loss function on the source domain dataset to generate an intermediate domain dataset similar to the target domain dataset; and step S23, merging the intermediate domain dataset and the source domain dataset to generate the hybrid domain dataset.

[0009] According to a specific embodiment of the present invention, the step of calculating the loss function of the source domain dataset to generate an intermediate domain dataset similar to the target domain dataset includes: calculating the loss functions of the source domain dataset and the target domain dataset to generate their corresponding migration images; first, converting the migration images through a generator, and then discriminating the converted migration images through a discriminator to generate the intermediate domain dataset.

[0010] According to a specific embodiment of the present invention, the steps of inputting the target domain dataset into the initial training model, extracting feature parameters from the target domain in the target domain dataset through the feature extraction network F, and then inputting the feature parameters into branch modules F1 and F2 to obtain the position and category information of the target domain and labeling it as the pseudo-label of the target domain, and simultaneously calculating the model optimization parameters, include: Step S41, extracting features from the target domain in the target domain dataset through the feature extraction network F to obtain the feature parameters corresponding to the target domain; Step S42, inputting the feature parameters into the branch modules F1 and F2 respectively; Step S43, calculating the loss functions of the branch modules F1 and F2 to obtain the model optimization parameters; Step S44, judging the position and category information of the target domain based on the position and category information obtained from the branch modules F1 and F2 through position thresholds and category thresholds, until all target domains in the target domain dataset are labeled.

[0011] According to a specific embodiment of the present invention, the model optimization parameter L is as follows:

[0012] L = L co +L1+L2

[0013]

[0014]

[0015]

[0016]

[0017]

[0018] in and It is to combine the convolutional blocks W in F1 and F2 con1 and W con2 The result after global averaging and channel compression of the weights; p and t represent the category and position information in the label, p1 and p1 represents the class and position obtained by the F1 branch module, while p2 and t2 represent the class and position obtained by the F2 branch module; classification loss L cls1 and L cls2 The regression loss L is calculated using the cross-entropy loss function. reg1 and L reg2 It is calculated using the smooth L1 function.

[0019] According to a specific embodiment of the present invention, step S44, which involves obtaining the location information and category information of the target domain based on the branch modules F1 and F2, and judging the location information and category information of the target domain by using location thresholds and category thresholds until all target domains in the target domain dataset are labeled, includes: step S441, where the branch modules F1 and F2 extract the feature parameters to generate prediction boxes K1 and K2, as well as location information I1 and category information C1 corresponding to K1, and location information I2 and category information C2 corresponding to K2; step S442, first screening: if the intersection-union ratio of the prediction boxes K1 and K2 with the preset ground truth boxes is greater than the location threshold, then a second screening is performed; step S443, second screening: when the category information C1 = C2, and the minimum value of the confidence scores of the category information C1 and C2 is greater than the category threshold, the maximum value of the location information I1 and I2 and the category information are used as pseudo-labels of the target domain.

[0020] A self-training system based on a hybrid domain and co-trained cross-domain detection model includes: an information acquisition module for acquiring source domain datasets and target domain datasets; an information extraction module for inputting the source domain dataset and the target domain dataset into a CycleGAN network for training to obtain a hybrid domain dataset; a first information training module for inputting the hybrid domain dataset into a co-training model for training to obtain an initial training model; a second information training module for inputting the target domain dataset into the initial training model, extracting feature parameters from the target domain in the target domain dataset through a feature extraction network F, and then inputting the feature parameters into branch modules F1 and F2 to obtain the location and category information of the target domain and label it as a pseudo-label of the target domain, while also calculating model optimization parameters; and a third information training module for inputting the target domain dataset labeled with pseudo-labels, the model optimization parameters, and the hybrid domain dataset into the initial training model to obtain a cross-domain detection model based on hybrid domain and co-training.

[0021] A self-training device based on a hybrid domain and co-trained cross-domain detection model includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any of the methods described above.

[0022] A computer-readable medium having instructions stored thereon, the instructions being loaded by a processor and executing the method described in any of the preceding claims.

[0023] The technical advantage of this invention lies in proposing a self-training method based on a hybrid domain and a co-trained cross-domain detection model. By effectively combining the source domain and the intermediate domain to form a hybrid domain, the technical problem of reduced image detection accuracy caused by domain differences can be effectively solved. The hybrid domain directly doubles the number of samples, enabling further improvement in network performance. After feature extraction from the image, the co-training structure in the model can observe the feature map from a dual-viewpoint, assigning higher-quality pseudo-labels to the target domain. The model is optimized by filtering target domain datasets with higher-quality pseudo-labels through preset position and category thresholds. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a self-training method for a cross-domain detection model based on hybrid domain and collaborative training, provided in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of the initial training model in a specific embodiment of the present invention;

[0026] Figure 3 This is a flowchart illustrating a self-training system based on a hybrid domain and collaborative training cross-domain detection model, according to an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of a self-training device based on a hybrid domain and collaborative training cross-domain detection model, provided in an embodiment of the present invention. Detailed Implementation

[0028] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0029] Please see Figure 1-4 It should be noted that the illustrations provided in this embodiment are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0030] This application embodiment uses a CycleGAN network to generate a hybrid domain to train a co-trained model, thereby obtaining a cross-domain detection model based on the hybrid domain and co-training. CycleGAN (Cycle Generative Adversarial Network) can help us achieve mutual conversion between images.

[0031] The self-training method based on the hybrid domain and collaborative training cross-domain detection model is applied to one or more electronic devices. The electronic device is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0032] The electronic device can be any electronic product that can interact with the user, such as a personal computer, tablet computer, smartphone, personal digital assistant (PDA), game console, interactive network television (IPTV), smart wearable device, etc.

[0033] The electronic device may also include network devices and / or user devices. The network devices include, but are not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers.

[0034] The networks in which the electronic devices are located include, but are not limited to, the Internet, wide area networks, metropolitan area networks, local area networks, and virtual private networks (VPNs).

[0035] Example 1

[0036] Please see Figure 1 The diagram illustrates the self-training method flow. A self-training method based on a hybrid domain and collaborative training cross-domain detection model includes:

[0037] Step S10: Collect the source domain dataset and the target domain dataset.

[0038] The source domain dataset includes the annotation information of the target objects, while the target domain dataset does not include the annotation information of the target objects.

[0039] Step S11: Normalize the size of the source domain dataset and the target domain dataset to obtain the processed source domain dataset. Target domain dataset Where, x i Let y represent the i-th image. i N represents the corresponding label. s N represents the number of samples in the source domain dataset. t This indicates the number of samples in the target domain dataset.

[0040] Step S20, the step of inputting the source domain dataset and the target domain dataset into the CycleGAN network to obtain a hybrid domain dataset includes:

[0041] Step S21: Input the source domain dataset and the target domain dataset into the CycleGAN network.

[0042] Step S22: Calculate the loss function for the source domain dataset to generate an intermediate domain dataset similar to the target domain dataset.

[0043] Step S221: Calculate the loss function of the source domain dataset and the target domain dataset, and generate the corresponding transfer image.

[0044] Step S222: First, the transfer image is transformed by the generator, and then the transformed transfer image is discriminated by the discriminator to generate an intermediate domain dataset.

[0045] Step S23: Mix the intermediate domain dataset and the source domain dataset to generate a hybrid domain dataset.

[0046] Recurrent Generative Adversarial Networks (CycleGANs) are powerful computer algorithms with the potential to improve the digital ecosystem by transforming information from one representation to another. For example, given an image, they can blur it, colorize it (if it was originally black and white), improve its sharpness, or fill in missing whitespace.

[0047] They are more powerful than traditional design / production / writing software. Because CycleGANs are machine learning algorithms, they can, in principle, learn to achieve any desired transformation. In contrast, traditional transformation software (such as Photoshop) is typically coded and developed to perform specific tasks. Moreover, CycleGANs can achieve higher performance from existing software because they learn from data and improve their effectiveness as more data is collected.

[0048] Preferably, in this embodiment, a CycleGAN network is used to transform the target to obtain a hybrid domain dataset.

[0049] Step S30: Input the mixed domain dataset into the collaborative training model for training to obtain the initial training model.

[0050] Step S40 involves inputting the target domain dataset into the initial training model, extracting feature parameters from the target domain in the target domain dataset using the feature extraction network F, and then inputting the feature parameters into branch modules F1 and F2 to obtain the location and category information of the target domain and label it as the pseudo-label of the target domain. Simultaneously, the steps of calculating model optimization parameters include:

[0051] Step S41: Extract features from the target domain in the target domain dataset using the feature extraction network F to obtain the feature parameters corresponding to the target domain;

[0052] In one specific embodiment, such as Figure 2 As shown, the feature extraction network F has 13 convolutional layers (i.e., kernel size of 3*3, edge padding of 1, stride of 1), 13 activation function layers, and 4 pooling layers (i.e., kernel size of 2*2, edge padding of 0, stride of 2).

[0053] Step S42: Input the feature parameters into branch modules F1 and F2 respectively;

[0054] In one specific embodiment, such as Figure 2 As shown, conv1 and conv2 in F1 and F2 are residual network blocks used to give the two residual blocks different weights and biases, thus resulting in different parameters calculated by F1 and F2. The RPN includes a conv layer (i.e., a 1*1 kernel with a stride of 1), a reshape layer (to obtain prediction boxes of different sizes), and a softmax classification layer.

[0055] Step S43: Calculate the loss functions of branch modules F1 and F2 to obtain the model optimization parameters.

[0056] The model optimization parameter L is as follows:

[0057] L = L co +L1+L2

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] in and It is to combine the convolutional blocks W in F1 and F2 con1 and W con2 The result after global averaging and channel compression of the weights; p and t represent the category and position information in the label, p1 and p1 represents the class and position obtained by the F1 branch module, while p2 and t2 represent the class and position obtained by the F2 branch module; classification loss L cls1 and L cls2 The regression loss L is calculated using the cross-entropy loss function. reg1 and L reg2 It is calculated using the smooth L1 function.

[0064] Neural network learning uses a metric to represent the current state, and then uses this metric as a benchmark to find the optimal weight parameters. The metric used in neural network learning is called the loss function. This loss function can be any function, but commonly uses mean squared error or cross-entropy error. The loss function is an indicator of the poor performance of the neural network, that is, the extent to which the current neural network does not fit the supervised data well, or is inconsistent with it.

[0065] The purpose of the loss function is to calculate the difference between the forward computation result of the neural network in each iteration and the true value, thereby guiding the next step of training in the right direction.

[0066] Step S44: Obtain the location information and category information of the target domain according to the branch modules F1 and F2. Judge the location information and category information of the target domain by the location threshold and category threshold. Use the location information and category information that meet the conditions as the pseudo-label of the corresponding target domain. If the conditions are not met, output the model optimization parameters, optimize the initial training model, and jump to step S41 until all target domains in the target domain dataset are labeled with pseudo-labels.

[0067] Step S441: The feature parameters are extracted through the reshape layer in the branch modules F1 and F2. The entire feature parameters are traversed by a sliding window to generate two different prediction boxes K1 and K2. At the same time, the position information I1 and I2 and the category information C1 and C2 of the prediction boxes K1 and K2 are obtained.

[0068] Step S442, First filtering: When IoU>∈ loc When the intersection-union ratios (IUUs) of predicted bounding boxes K1 and K2 with the preset ground truth bounding boxes are both greater than the position threshold, the condition is met. A second round of filtering is then performed.

[0069] Step S443, Second Filtering: Let y1 and y2 represent the confidence levels of C1 and C2, respectively. A decision is made if and only if C1 = C2 and min(y1, y2) > ∈ [0, 0]. cls When the location information I1 and I2 are equal, the larger value is used as the location information of the corresponding target domain. If the category information is equal, one of them is selected. The location information and category information are used as pseudo-labels of the corresponding target domain.

[0070] Step S50: Input the target domain dataset and the hybrid domain dataset labeled with pseudo-labels into the current initial training model for training, and obtain a cross-domain detection model based on hybrid domain and collaborative training.

[0071] The self-training method of the collaborative training model described in this embodiment feeds both the target domain dataset with pseudo-labels and the mixed domain dataset into the model for training and parameter adjustment. It is important to note that the model initially selects only a portion of the target domain images for pseudo-labeling, rather than the entire target domain, to avoid an excessive amount of erroneous pseudo-label data that could lead to negative transfer of the model due to an excessively high proportion of erroneous data in the labeling process.

[0072] The cross-domain detection model based on hybrid domain and co-training provided in this embodiment was tested on several mainstream publicly available domain adaptation datasets. Extensive comparative experiments were conducted on these datasets, and the proposed model achieved better results compared to current mainstream domain adaptation models, demonstrating its high performance and robustness.

[0073] It should be noted that, in order to ensure data security, the data involved can be deployed on a blockchain to prevent the data from being maliciously tampered with.

[0074] It should be noted that the steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0075] Example 2

[0076] like Figure 3 As shown, this embodiment discloses a self-training system based on a hybrid domain and co-trained cross-domain detection model, including:

[0077] Information acquisition module 10 is used to acquire source domain datasets and target domain datasets;

[0078] The information extraction module 20 is used to input the source domain dataset and the target domain dataset into the CycleGAN network for training to obtain a hybrid domain dataset;

[0079] The first information training module 30 is used to input the hybrid domain dataset into the collaborative training model for training to obtain an initial training model;

[0080] The second information training module 40 is used to input the target domain dataset into the initial training model, extract feature parameters of the target domain in the target domain dataset through the feature extraction network F, and then input the feature parameters into the branch modules F1 and F2 to obtain the position and category information of the target domain and mark it as the pseudo label of the target domain. At the same time, the model optimization parameters are also calculated.

[0081] The third information training module 50 is used to input the target domain dataset labeled with pseudo-labels, model optimization parameters, and hybrid domain dataset into the initial training model for training, so as to obtain a cross-domain detection model based on hybrid domain and collaborative training.

[0082] Example 3

[0083] like Figure 4 As shown, this embodiment discloses a self-training device based on a hybrid domain and collaborative training cross-domain detection model, including a memory 2, a processor 1, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any of the methods described above.

[0084] The memory includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory can be an external storage device of the electronic device, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory can include both internal and external storage units of the electronic device. The memory can be used not only to store application software and various types of data installed on the electronic device, but also to temporarily store data that has been output or will be output.

[0085] In some embodiments, a processor may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control unit of the electronic device, connecting various components of the device via various interfaces and lines. It executes programs or modules stored in the memory and calls data stored in the memory to perform various functions and process data within the electronic device.

[0086] The processor executes the operating system of the electronic device and various installed applications. The processor executes the applications to implement the steps in the above embodiments of the lithium-ion battery cold solder joint detection method.

[0087] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0088] The integrated unit, implemented as a software functional module, can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute some functions of the lithium battery cold solder joint detection method of the various embodiments of the present invention.

[0089] In summary, the technical advantage of this invention lies in proposing a self-training method based on a hybrid domain and a co-trained cross-domain detection model. By effectively combining the source domain and the intermediate domain to form a hybrid domain, the technical problem of reduced image detection accuracy caused by domain differences can be effectively solved. The hybrid domain directly doubles the number of samples, enabling further improvement in network performance. After feature extraction from the image, the co-training structure in the model can observe the feature map from a dual-viewpoint, assigning higher-quality pseudo-labels to the target domain. The model is optimized by filtering target domain datasets with higher-quality pseudo-labels through preset position and category thresholds.

[0090] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0091] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A self-training method based on a hybrid domain and co-trained cross-domain detection model, characterized in that, include: Collect source domain datasets and target domain datasets; The source domain dataset and the target domain dataset are input into the CycleGAN network to obtain a hybrid domain dataset, and the steps include: Step S21: Input the source domain dataset and the target domain dataset into the CycleGAN network; Step S22: Calculate the loss function for the source domain dataset to generate an intermediate domain dataset similar to the target domain dataset; wherein, calculate the loss functions for the source domain dataset and the target domain dataset to generate their corresponding migration images; first, the migration images are transformed by a generator, and then the transformed migration images are discriminated by a discriminator to generate the intermediate domain dataset; Step S23: The intermediate domain dataset and the source domain dataset are merged to generate the hybrid domain dataset; The hybrid domain dataset is input into the collaborative training model for training to obtain the initial training model; The target domain dataset is input into the initial training model. Feature parameters are extracted from the target domain in the target domain dataset using the feature extraction network F. These feature parameters are then input into branch modules F1 and F2 to obtain the location and category information of the target domain and label it as a pseudo-label for the target domain. Simultaneously, model optimization parameters are calculated. The steps include: Step S41: Extract features from the target domain in the target domain dataset using the feature extraction network F to obtain the feature parameters corresponding to the target domain; Step S42: Input the feature parameters into the branch modules F1 and F2 respectively; Step S43: Calculate the loss functions of the branch modules F1 and F2 to obtain the model optimization parameters; Step S44: Obtain the location information and category information of the target domain according to the branch modules F1 and F2, and judge the location information and category information of the target domain by the location threshold and category threshold until all target domains in the target domain dataset are labeled. The target domain dataset labeled with pseudo-labels, model optimization parameters, and hybrid domain dataset are input into the initial training model to obtain a cross-domain detection model based on hybrid domain and co-training. The model optimization parameter L is as follows: L=L co +L1+L2 in and It is to combine the convolutional blocks W in F1 and F2 con1 and W con2 The result after global averaging and channel compression of the weights; p and t represent the category and position information in the label, p1 and p1 represents the class and position obtained by the F1 branch module, while p2 and t2 represent the class and position obtained by the F2 branch module; classification loss L cls1 and L cls2 The regression loss L is calculated using the cross-entropy loss function. reg1 and L reg2 It was calculated using the smooth L1 function.

2. The self-training method according to claim 1, characterized in that, The process of collecting the source domain dataset and the target domain dataset also includes: performing size normalization processing on the source domain dataset and the target domain dataset.

3. The self-training method according to claim 1, characterized in that, Step S44, which involves obtaining the location and category information of the target domain based on the branch modules F1 and F2, and judging the location and category information of the target domain using location and category thresholds until all target domains in the target domain dataset are labeled, includes the following steps: In step S441, the branch modules F1 and F2 extract the feature parameters to generate prediction boxes K1 and K2, as well as the location information I1 and category information C1 corresponding to K1, and the location information I2 and category information C2 corresponding to K2. Step S442, First filtering: If the intersection-union ratios of the predicted boxes K1 and K2 with the preset real boxes are both greater than the position threshold, then a second filtering is performed; Step S443, second filtering: When the category information C1 = C2, and the minimum value of the confidence of the category information C1 and C2 is greater than the category threshold, the maximum value of the location information I1 and I2 and the category information are used as pseudo-labels of the target domain.

4. A self-training system based on a hybrid domain and collaborative training cross-domain detection model, characterized in that, include: The information acquisition module is used to collect source domain datasets and target domain datasets; The information extraction module is used to input the source domain dataset and the target domain dataset into the CycleGAN network for training to obtain a hybrid domain dataset, and the steps include: Step S21: Input the source domain dataset and the target domain dataset into the CycleGAN network; Step S22: Calculate the loss function for the source domain dataset to generate an intermediate domain dataset similar to the target domain dataset; wherein, calculate the loss functions for the source domain dataset and the target domain dataset to generate their corresponding migration images; first, the migration images are transformed by a generator, and then the transformed migration images are discriminated by a discriminator to generate the intermediate domain dataset; Step S23: The intermediate domain dataset and the source domain dataset are merged to generate the hybrid domain dataset; The first information training module is used to input the hybrid domain dataset into the collaborative training model for training to obtain an initial training model; The second information training module is used to input the target domain dataset into the initial training model, extract feature parameters from the target domain in the target domain dataset through the feature extraction network F, and then input the feature parameters into branch modules F1 and F2 to obtain the location and category information of the target domain and label it as the pseudo-label of the target domain. Simultaneously, it also calculates the model optimization parameters, and the steps include: Step S41: Extract features from the target domain in the target domain dataset using the feature extraction network F to obtain the feature parameters corresponding to the target domain; Step S42: Input the feature parameters into the branch modules F1 and F2 respectively; Step S43: Calculate the loss functions of the branch modules F1 and F2 to obtain the model optimization parameters; Step S44: Obtain the location information and category information of the target domain according to the branch modules F1 and F2, and judge the location information and category information of the target domain by the location threshold and category threshold until all target domains in the target domain dataset are labeled. The third information training module is used to input the target domain dataset labeled with pseudo-labels, model optimization parameters, and hybrid domain dataset into the initial training model to obtain a cross-domain detection model based on hybrid domain and collaborative training. The model optimization parameter L is as follows: L=L co +L1+L2 in and It is to combine the convolutional blocks W in F1 and F2 con1 and W con2 The result after global averaging and channel compression of the weights; p and t represent the category and position information in the label, p1 and p1 represents the class and position obtained by the F1 branch module, while p2 and t2 represent the class and position obtained by the F2 branch module; classification loss L cls1 and L cls2 The regression loss L is calculated using the cross-entropy loss function. reg1 and L reg2 It is calculated using the smooth L1 function.

5. A self-training device based on a hybrid domain and collaborative training cross-domain detection model, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 3.

6. A computer-readable medium, characterized in that, It stores instructions that are loaded by a processor and executed as described in any one of claims 1 to 3.

Citation Information

Patent Citations

  • Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption

    CN114429152A

  • CT image generation method and apparatus, computer device, and computer-readable storage medium

    US20210251590A1