Passive domain adaptive object detection method and system based on decoupled teacher model
Patent Information
- Application Number
- CN202511175891.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-21
- Publication Date
- 2026-10-30
- Estimated Expiration
- 2045-08-21
AI Technical Summary
[0004]为了解决上述问题，本发明提出一种基于解耦教师模型的无源域自适应目标检测方法及系统，有效解决了领域自适应过程中历史知识主导与新知识获取之间的动态失衡问题
1、本发明提出了解耦教师框架，通过学生模型梯度敏感度分析构建参数更新反馈环路，有效解决了领域自适应过程中历史知识主导与新知识获取之间的动态失衡问题。
Smart Images
Figure CN121072648B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, and in particular to a passive domain adaptive target detection method and system based on a decoupled teacher model. Background Technology
[0002] Passive Domain Adaptive Object Detection (SF-DAOD) protects data privacy by restricting access to source data during the adaptation process, while transferring knowledge gained from the labeled source domain to the unlabeled target domain. Compared to traditional domain adaptive object detection, the passive domain adaptation method faces greater challenges due to the lack of guidance from source domain data. Existing methods mainly utilize the mean teacher framework for self-training in the target domain. The exponential moving average (EMA) mechanism in the mean teacher framework maintains a stable teacher model by averaging student weights during the training steps. High-quality pseudo-labels are generated from the unlabeled data using the stable teacher model, and the student model adapts to the target domain using these pseudo-labels.
[0003] However, the inherent lag in the teacher model's response to new knowledge hinders the student's rapid adaptation to changes in the target domain. In traditional supervised learning, a stable teacher model can withstand training perturbations. But for passive domain adaptive models, this advantage becomes a key constraint. A teacher model that relies too heavily on past knowledge will hinder the student model, while a teacher model that relies too heavily on current knowledge will be unable to withstand training perturbations. Summary of the Invention
[0004] To address the aforementioned issues, this invention proposes a passive domain adaptive target detection method and system based on a decoupled teacher model, which effectively solves the dynamic imbalance between historical knowledge dominance and new knowledge acquisition during domain adaptation.
[0005] The technical solution adopted in this invention is as follows: A passive domain adaptive target detection method based on a decoupled teacher model includes: Passive domain adaptive object detection training is performed using a teacher-student model network architecture, and the pseudo-labels generated by the teacher model and student model are combined to obtain integrated pseudo-labels. The integrated pseudo-labels are used for backpropagation to update the parameters of the student model, and the parameters of the teacher model are updated by combining parameter importance and asynchronous exponential moving average mechanism. The trained teacher-student model network architecture was used for passive domain adaptive object detection.
[0006] Furthermore, when combining the pseudo-labels generated by the teacher model and the student model, a non-maximum suppression method is used to avoid the occurrence of two labels for the same target.
[0007] Furthermore, the process of combining the pseudo-labels generated by the teacher model and the student model to obtain integrated pseudo-labels includes:
[0008] in, This indicates the integration of pseudo-tags. This indicates a non-maximum suppression operation. Indicates a connection operation. This refers to the pseudo-labels generated by the teacher model. This represents the pseudo-labels generated by the student model.
[0009] Furthermore, the method of updating the parameters of the teacher model by combining parameter importance and asynchronous exponential moving average mechanism includes: sorting the parameters by importance based on the gradient magnitude, grouping the parameters according to the parameter importance sorting, and asynchronously updating the model parameters according to the parameter grouping.
[0010] Furthermore, the parameter importance ranking based on gradient magnitude includes: using the gradient derived from backpropagation as an intrinsic indicator of the parameter's contribution to loss minimization, with the optimization objective being to minimize the detection loss between the teacher model prediction and the aggregated pseudo-labels in the target domain.
[0011] Furthermore, the method for calculating the importance of the parameter includes:
[0012] in, Represents the parameters in the teacher model Importance, || represents absolute value, Indicates the prediction results based on the teacher model. and integrated pseudo tags Calculation of target detection loss.
[0013] Furthermore, the parameter grouping based on parameter importance includes: selecting the parameter with the highest parameter importance as the target response parameter. Other parameters serve as stability parameters. .
[0014] Furthermore, the asynchronous update of model parameters based on parameter grouping includes: for the target response parameters Assign a smoothing coefficient below the first threshold to accommodate distribution drift in the target domain; to maintain stability parameters A smoothing coefficient higher than the second threshold is assigned to ensure the integrity of historical knowledge; wherein the first threshold is lower than the second threshold.
[0015] Furthermore, the parameter update strategies for the teacher model and the student model include:
[0016] in, The first teacher model One parameter, The student model represents the first One parameter, Target response parameters The smoothing coefficient, It is a stability parameter The smoothing coefficient.
[0017] A passive domain adaptive target detection system based on a decoupled teacher model includes: An integrated pseudo-label generation module is configured to perform passive domain adaptive object detection training through a teacher-student model network architecture, and combine the pseudo-labels generated by the teacher model and the student model to obtain integrated pseudo-labels; The parameter update module is configured to use integrated pseudo-labels for backpropagation to update the parameters of the student model, and to update the parameters of the teacher model by combining parameter importance and asynchronous exponential moving average mechanism. The object detection module is configured to use a trained teacher-student model network architecture for passive domain adaptive object detection.
[0018] The beneficial effects of this invention are as follows: 1. This invention proposes a decoupled teacher framework, which constructs a parameter update feedback loop through gradient sensitivity analysis of the student model, effectively solving the dynamic imbalance between historical knowledge dominance and new knowledge acquisition in the process of domain adaptation.
[0019] 2. This invention proposes a knowledge fusion mechanism to facilitate the generation of comprehensive labels that include both past and current knowledge. This mechanism utilizes the current knowledge of the student model to assist the teacher model in generating pseudo-labels based on its past knowledge.
[0020] 3. This invention proposes an asynchronous exponential moving average mechanism, redefining the parameter update paradigm of the teacher model. This mechanism breaks through the static smoothing strategy of the traditional mean-based teacher framework, innovatively introducing a parameter importance measurement method based on gradient magnitude, thus realizing dynamic grouping and differentiated updates of network parameters.
[0021] 4. Experiments conducted on multiple migration datasets demonstrate that the present invention achieves superior performance, outperforming state-of-the-art methods to date. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of a passive domain adaptive target detection method based on a decoupled teacher model, according to Embodiment 2 of the present invention. Detailed Implementation
[0023] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0024] Example 1 like Figure 1 As shown, this embodiment provides a passive domain adaptive target detection method based on a decoupled teacher model, including: Passive domain adaptive object detection training is performed using a teacher-student model network architecture, and the pseudo-labels generated by the teacher model and student model are combined to obtain integrated pseudo-labels. The integrated pseudo-labels are used for backpropagation to update the parameters of the student model, and the parameters of the teacher model are updated by combining parameter importance and asynchronous exponential moving average mechanism. Finally, the trained teacher-student model network architecture was used for passive domain adaptive object detection.
[0025] Preferably, when combining the pseudo-labels generated by the teacher model and the student model, a non-maximum suppression method is used to avoid the occurrence of two labels for the same target.
[0026] Preferably, the parameters of the teacher model are updated by combining parameter importance and an asynchronous exponential moving average mechanism, including: ranking the parameters based on the gradient magnitude, grouping the parameters according to the ranking, and updating the model parameters asynchronously according to the parameter groups.
[0027] More preferably, the parameter importance is ranked based on the gradient magnitude, including: using the gradient derived from backpropagation as an intrinsic indicator of the parameter's contribution to minimizing the loss, with the optimization objective being to minimize the detection loss between the teacher model prediction and the aggregated pseudo-labels in the target domain.
[0028] More preferably, the parameters are grouped according to their importance, including selecting the parameter with the highest importance as the target response parameter. Other parameters serve as stability parameters. .
[0029] More preferably, the model parameters are updated asynchronously according to parameter grouping, including: for the target response parameters Assign a smoothing coefficient below the first threshold to accommodate distribution drift in the target domain; to maintain stability parameters A smoothing coefficient higher than the second threshold is assigned to ensure the integrity of historical knowledge; wherein the first threshold is lower than the second threshold.
[0030] It should be noted that this method uses an asynchronous exponential moving average (EMA) mechanism to balance the stability and adaptability of the model. Unlike traditional EMA, which updates all parameters uniformly, asynchronous EMA dynamically classifies parameters based on their importance to the target domain and assigns corresponding smoothing coefficients.
[0031] Specifically, to acquire comprehensive domain knowledge, pseudo-labels from both teachers and students are integrated. The importance of each parameter is quantified by the magnitude of the backpropagation gradient of the teacher's loss function, which measures the difference between the teacher's predictions and the integrated pseudo-labels. Parameters with larger gradients (indicating a greater impact on reducing prediction error) are prioritized during optimization. Note that the teacher gradient guides parameter grouping but is not used in the gradient descent algorithm.
[0032] Furthermore, this method identifies a sparse subset of key parameters, termed the target response parameters, which exhibit significant importance for domain adaptation. The remaining parameters are categorized as stability-preserving parameters to maintain the integrity of temporal knowledge. To prioritize the rapid adaptation of these key parameters, a lower smoothing coefficient is assigned to the target response parameters, enabling them to be updated quickly in response to changes in the target domain. In contrast, stability-preserving parameters are controlled by a higher smoothing coefficient, which introduces inertia into their updates and prevents catastrophic forgetting. This grouped update method enhances the adaptability of the teacher model while maintaining overall stability. By adjusting the smoothing coefficient according to parameter importance, it can rapidly adapt to new knowledge in emerging domains without sacrificing the stability of the learned knowledge structure.
[0033] Example 2 This embodiment is based on embodiment 1: like Figure 1 As shown, this embodiment provides a passive domain adaptive target detection method based on a decoupled teacher model, which dynamically groups the teacher model parameters into target response parameters and stability-preserving parameters. A lower smoothing coefficient is assigned to the target response parameters to achieve rapid adaptation, while a higher smoothing coefficient is assigned to the stability-preserving parameters to maintain historical knowledge. Specific details are as follows.
[0034] I. Knowledge Integration Combining pseudo-labels generated by teacher and student models enhances domain knowledge acquisition. This fusion compensates for the differences arising from their varying learning strategies: teacher models focus on maintaining stable knowledge through incremental updates, while student models rapidly adapt to new patterns in the target domain. By combining the teacher's long-term knowledge preservation with the student's short-term pattern adaptation, this combined approach reduces individual model bias and strengthens the overall learning process.
[0035] Specifically, the historical consistency of the teacher model balances the tendency of the student model to overfit transient trends, while the responsiveness of the student model prevents the teacher model from becoming rigid. The resulting integration of consistent historical understanding and flexible pattern recognition facilitates more reliable and robust domain knowledge acquisition, ultimately bridging the gap between generalization and domain-specific adaptation.
[0036] Preferably, considering that the teacher model and the student model detect a large number of identical targets, non-maximum suppression is used to avoid the same target having two labels. This is given by pseudo-labels from the teacher model. (using threshold) (Filtered predictions) and pseudo-labels from the student model (using threshold) (Filtering and prediction), the integration process can be described as follows:
[0037] in, This indicates the integration of pseudo-tags. This indicates a non-maximum suppression operation. This indicates a join operation.
[0038] II. Asynchronous Exponential Moving Average Mechanism This method proposes an asynchronous exponential moving average (AEMA) mechanism, which achieves asynchronous teacher parameter updates through parameter grouping and dual-rate updates, including two steps: (1) gradient-aware parameter importance ranking; (2) grouped updates.
[0039] The parameter importance ranking requires defining a measurable surrogate to quantify the importance of each parameter during domain adaptation. This is achieved by analyzing the gradient descent algorithm: the magnitude of the gradient serves as an intrinsic indicator of a parameter's contribution to loss minimization. The optimization objective is set to minimize the detection loss between the teacher model's predictions and the aggregated pseudo-labels in the target domain. The importance index is derived from the gradients derived from backpropagation; larger gradient magnitudes correspond to higher domain adaptation priority and require priority updates to facilitate effective knowledge learning in the target domain. Conversely, parameters with smaller gradients are identified as stability-preserving components, which play a crucial role in maintaining the consistency of historical knowledge.
[0040]
[0041] in, Represents the parameters in the teacher model Importance, || represents absolute value, Indicates the prediction results based on the teacher model. and integrated pseudo tags Calculation of target detection loss.
[0042] Preferably, based on the teacher parameters' ability to quickly adapt to new domain knowledge, teacher parameters can be divided into two categories: target response parameters and stability-preserving parameters. Inspired by network pruning, only a few parameters are essential for network learning. Given that each parameter... The importance of which The number of teacher parameters is used to select the parameter with the highest importance as the target response parameter. Other parameters serve as stability parameters. .
[0043] Preferably, in grouped asynchronous updates, a lower smoothing coefficient is assigned to the target response parameter to quickly adapt to the distributional drift of the target domain; a higher smoothing coefficient is assigned to the stability-preserving parameter to ensure the integrity of historical knowledge. This dual-coefficient strategy reflects their different roles: the target response parameter prioritizes agility, dynamically matching the evolving target pattern; while the stability-preserving parameter emphasizes inertia, mitigating sudden deviations from consolidated historical knowledge. By decoupling the update rate based on parameter roles, this method balances the adaptability of plasticity with the stability of long-term consistency.
[0044] Finally, the updates for each parameter are described below:
[0045] in, The first teacher model One parameter, The student model represents the first One parameter, Target response parameters The smoothing coefficient, It is a stability parameter The smoothing coefficient.
[0046] III. Training Objectives Masked Image Consistency (MIC) is used during the training phase to facilitate learning the contextual relationships of unlabeled data. Given a strongly augmented image... mask image Depend on Generate, where It is a patch mask sampled from a uniform distribution. This represents element-wise multiplication. The overall training objective can be stated as follows:
[0047] in, and These represent the predictions of the student model. and .
[0048] IV. Implementation Details Specifically, this embodiment uses Deformable DETR as the base detector. For all domain adaptation tasks, the batch size is fixed at 8. The teacher model is updated using an asynchronous exponential moving average (AEMA) mechanism, where the smoothing coefficient is set to... , (For Cityscapes → BDD100K, , The top 10% of parameters are selected as the target response parameters. The threshold for pseudo-labels is... , The student model was optimized using the Adam optimizer with an initial learning rate of 2e-4. The MIC module was used, with a patch size of 64 and a mask ratio of 0.5. Mean Average Precision (mAP) with a threshold of 0.5 was used as the evaluation metric. Experiments on several typical transfer learning tasks validated the model, demonstrating that this approach achieves optimal performance.
[0049] Experimental verification has been conducted on several typical transfer tasks, and the experimental results are shown in Tables 1 to 3.
[0050] Table 1: Comparison of Experimental Results for the Cityscape → FoggyCityscape Migration Task
[0051] As shown in Table 1, compared with DRU using the same detector framework, our method improves performance on difficult categories (such as "train" and "motorcycle"), where more teacher-generated pseudo-labels are generated and better adapted. Therefore, our method achieves state-of-the-art performance.
[0052] Table 2: Comparison of Experimental Results for the Cityscape→BDD100K Migration Task
[0053] Table 2 shows the experimental results for Cityscapes → BDD100K. This method improves mAP by 3% compared to the previous state-of-the-art method and achieves the best performance in all categories.
[0054] Table 3: Comparison of Experimental Results for the Sim10K→Cityscape Transfer Task
[0055] As shown in Table 3, the proposed framework demonstrates superior performance compared to existing methods, achieving at least a 2% improvement in mAP compared to the DefDETR-based DAOD method.
[0056] In summary, as shown in Tables 1 to 3, the passive domain adaptive target detection method based on the decoupled teacher model in this embodiment achieves the best performance.
[0057] Example 3 This embodiment provides a passive domain adaptive target detection system based on a decoupled teacher model, including: An integrated pseudo-label generation module is configured to perform passive domain adaptive object detection training through a teacher-student model network architecture, and combine the pseudo-labels generated by the teacher model and the student model to obtain integrated pseudo-labels; The parameter update module is configured to use integrated pseudo-labels for backpropagation to update the parameters of the student model, and to update the parameters of the teacher model by combining parameter importance and asynchronous exponential moving average mechanism. The object detection module is configured to use a trained teacher-student model network architecture for passive domain adaptive object detection.
[0058] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the passive domain adaptive target detection method based on a decoupled teacher model according to Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.
[0059] Example 5 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the passive domain adaptive target detection method based on a decoupled teacher model according to Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the storage medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electrical carrier signals and telecommunication signals.
[0060] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
[0061] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. Furthermore, the terms "first," "second," and "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.
Claims
1. A passive domain adaptive target detection method based on a decoupled teacher model, characterized in that, include: Passive domain adaptive image target detection training is performed using a teacher-student model network architecture, and the pseudo-labels generated by the teacher model and student model are combined to obtain integrated pseudo-labels. The integrated pseudo-labels are used for backpropagation to update the parameters of the student model, and the parameters of the teacher model are updated by combining parameter importance and asynchronous exponential moving average mechanism. The trained teacher-student model network architecture was used for passive domain adaptive image object detection. When combining the pseudo-labels generated by the teacher model and the student model, the non-maximum suppression method is used to avoid the occurrence of two labels for the same target; The process of combining the pseudo-labels generated by the teacher model and the student model to obtain integrated pseudo-labels includes: in, This indicates the integration of pseudo-tags. This indicates a non-maximum suppression operation. Indicates a connection operation. This refers to the pseudo-labels generated by the teacher model. This refers to the pseudo-labels generated by the student model; The method of updating the parameters of the teacher model by combining parameter importance and asynchronous exponential moving average mechanism includes: sorting the parameters by importance based on the gradient magnitude, grouping the parameters according to the parameter importance sorting, and asynchronously updating the model parameters according to the parameter grouping.
2. The passive domain adaptive target detection method based on a decoupled teacher model according to claim 1, characterized in that, The parameter importance ranking based on gradient magnitude includes: using the gradient derived from backpropagation as an intrinsic indicator of the parameter's contribution to loss minimization, with the optimization objective being to minimize the detection loss between the teacher model prediction and the aggregated pseudo-labels in the target domain.
3. The passive domain adaptive target detection method based on a decoupled teacher model according to claim 2, characterized in that, The methods for calculating the importance of the parameters include: in, Represents the parameters in the teacher model Importance, || represents absolute value, Indicates the prediction results based on the teacher model. and integrated pseudo tags Calculation of target detection loss.
4. The passive domain adaptive target detection method based on a decoupled teacher model according to claim 1, characterized in that, The parameter grouping based on parameter importance includes: selecting the parameter with the highest importance as the target response parameter. Other parameters serve as stability parameters. .
5. The passive domain adaptive target detection method based on a decoupled teacher model according to claim 4, characterized in that, The asynchronous update of model parameters based on parameter grouping includes: for the target response parameters Assign a smoothing coefficient below the first threshold to accommodate distribution drift in the target domain; to maintain stability parameters A smoothing coefficient higher than the second threshold is assigned to ensure the integrity of historical knowledge; wherein the first threshold is lower than the second threshold.
6. The passive domain adaptive target detection method based on a decoupled teacher model according to claim 5, characterized in that, The parameter update strategies for the teacher and student models include: in, The teacher model represents the first One parameter, The student model represents the first One parameter, Target response parameters The smoothness coefficient, It is a stability parameter The smoothing coefficient.
7. A passive domain adaptive target detection system based on a decoupled teacher model, characterized in that, include: An integrated pseudo-label generation module is configured to perform passive domain adaptive image target detection training through a teacher-student model network architecture, and combine the pseudo-labels generated by the teacher model and the student model to obtain integrated pseudo-labels; The parameter update module is configured to use integrated pseudo-labels for backpropagation to update the parameters of the student model, and to update the parameters of the teacher model by combining parameter importance and asynchronous exponential moving average mechanism. The object detection module is configured to use a trained teacher-student model network architecture for passive domain adaptive image object detection. When combining the pseudo-labels generated by the teacher model and the student model, the non-maximum suppression method is used to avoid the occurrence of two labels for the same target; The process of combining the pseudo-labels generated by the teacher model and the student model to obtain integrated pseudo-labels includes: in, This indicates the integration of pseudo-tags. This indicates a non-maximum suppression operation. Indicates a connection operation. This refers to the pseudo-labels generated by the teacher model. This refers to the pseudo-labels generated by the student model; The method of updating the parameters of the teacher model by combining parameter importance and asynchronous exponential moving average mechanism includes: sorting the parameters by importance based on the gradient magnitude, grouping the parameters according to the parameter importance sorting, and asynchronously updating the model parameters according to the parameter grouping.