A passive domain adaptive target detection method and electronic device based on pseudo-label confidence feedback
By introducing a pseudo-label confidence feedback mechanism into the Mean-Teacher architecture, the EMA decay rate and pseudo-label screening threshold are dynamically adjusted, which solves the problem of pseudo-label noise accumulation in passive domain adaptive tasks, realizes stable model transfer and efficient training in the target domain, and improves detection performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2026-04-20
- Publication Date
- 2026-05-26
AI Technical Summary
The fixed EMA strategy in the existing Mean-Teacher architecture leads to the accumulation of pseudo-label noise, which causes the model to degrade detection performance in the target domain, and lacks effective pseudo-label quality control in passive domain adaptation tasks.
By introducing a pseudo-label confidence feedback mechanism, dynamically adjusting the pseudo-label screening threshold and EMA decay rate, an adaptive mean teacher update strategy is constructed. The confidence signal is used to optimize the weight update of the teacher model, suppress the accumulation of pseudo-label noise, and improve the model's transfer stability in the target domain.
It effectively suppressed the accumulation of false label noise, improved the training stability and detection accuracy of the model in the target domain, enhanced cross-domain transfer efficiency, reduced training costs, and strengthened the model's generalization ability.
Smart Images
Figure CN122090211A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision, specifically to a passive domain adaptive target detection method and electronic device based on pseudo-label confidence feedback. Background Technology
[0002] Object detection is one of the core tasks in computer vision. In recent years, deep learning-based object detection algorithms (such as the YOLO series) have achieved great success on various benchmark datasets. However, in real-world applications, due to differences in shooting equipment, lighting and weather conditions (e.g., from sunny to foggy days), and data sources, there is often a significant domain shift between the training data (source domain) and the test data (target domain). This domain shift can lead to a substantial decrease in the detection performance of models trained on the source domain on the target domain.
[0003] To address this issue, unsupervised domain adaptation (UDA) techniques have been extensively studied. Traditional UDA methods typically require simultaneous access to labeled source domain data and unlabeled target domain data. However, due to reasons such as data privacy protection, data transfer costs, and storage limitations, obtaining source domain data is impractical in many real-world scenarios. Therefore, the passive domain adaptation (SFDA) task has emerged. In the SFDA task, the model adaptation process can only utilize a pre-trained source domain model and unlabeled target domain data.
[0004] In SFDA object detection tasks, current mainstream methods (such as SF-YOLO) typically employ a Mean-Teacher architecture combined with pseudo-labels for unsupervised self-training. In this architecture, the student model is trained using pseudo-labels generated by the teacher model, and the teacher model's weights are updated from the student model using an Exponential Moving Average (EMA). Existing Mean-Teacher methods have a significant drawback: the teacher model's weight updates utilize a fixed decay rate EMA strategy. In the early stages of adaptive training in the target domain, or when the model encounters difficult samples, the student model generates a large number of low-quality, noisy pseudo-labels. Since the EMA update rate is fixed, the teacher model inevitably passively absorbs this noisy weight information. As training progresses, this noise continuously amplifies between the student and teacher models, generating severe "confirmation bias," ultimately leading to model collapse or a sharp decline in detection performance. Summary of the Invention
[0005] The purpose of this invention is to provide a passive domain adaptive target detection method and electronic device based on pseudo-label confidence feedback, which aims to solve the "confirmation bias" problem caused by the fixed EMA strategy in the existing Mean-Teacher architecture, effectively suppress the accumulation of pseudo-label noise, and achieve smooth, fast and stable model migration to the target domain.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A passive domain adaptive target detection method based on pseudo-label confidence feedback includes the following steps:
[0008] S1. Obtain the pre-trained source domain model and unlabeled target domain data: Obtain unlabeled target domain data that contains only image data and no labeling information, and use the model that has been trained on the source domain as the pre-trained source domain model on the target domain.
[0009] S2. Construct the mean teacher self-training architecture: Based on the pre-trained source domain model obtained in step S1, construct the mean teacher self-training architecture, which includes student model and teacher model. The student model and teacher model have the same network structure, and the initial weights are assigned by the pre-trained source domain model.
[0010] S3. Generate pseudo-labels and calculate confidence feedback signals: Input the image data of the target domain obtained in step S1 into the teacher model constructed in step S2 to generate target prediction boxes and their corresponding confidence scores; dynamically adjust the pseudo-label selection threshold according to the current training state, and obtain pseudo-labels based on the threshold selection; at the same time, calculate the confidence statistics of pseudo-labels in the current training batch, and use them as feedback signals for dynamic adjustment.
[0011] S4. Adaptive Dynamic Mean Teacher Update Strategy: The student model constructed in step S2 is used to train the target domain data obtained in step S1. During the weight update phase, the momentum decay rate parameter of the exponential moving average (EMA) is dynamically calculated and adjusted based on the confidence feedback signal obtained in step S3. The pseudo-label screening threshold and momentum decay rate parameter are adjusted when the model validation performance changes. The updated momentum decay rate is used to update the weights of the student model to the teacher model.
[0012] Preferably, in step S2, the network structure in the mean teacher self-training architecture is based on the SF-YOLO object detection model; the weights of the student model are updated through backpropagation using the gradient descent algorithm, while the weights of the teacher model do not participate in backpropagation and are only updated by receiving the weights of the student model through exponential moving average (EMA).
[0013] Preferably, the specific process of generating pseudo-labels and calculating confidence feedback signals in step S3 includes:
[0014] S31. Pseudo-label generation and preliminary screening: Input the target domain image into the teacher model to obtain the predicted bounding box, class probability, and target score; for the j-th image in the current training batch, select the set of predicted bounding boxes whose confidence level output by the teacher model is greater than the pseudo-label confidence screening threshold τ. Determine the pseudo-label filtering threshold for the current batch based on the current training status. The prediction results are then filtered using a pseudo-label selection threshold, retaining high-confidence prediction boxes as pseudo-labels for the current batch. ;
[0015] S32. Extract the confidence feedback signal: Calculate the average confidence score of all selected pseudo-labels in the current batch, as a feedback signal reflecting the current performance of the teacher model. And calculate the average confidence level of all valid pseudo-labels in the current batch. This is used as a confidence index to measure the current model's adaptability to the target domain data and the quality of the pseudo-labels;
[0016] S33. Feedback signal smoothing: The feedback signal is smoothed using a historical moving average to obtain a stable feedback index. .
[0017] Preferably, the specific process of extracting the confidence feedback signal in step S32 includes:
[0018] Feedback signal Represented as:
[0019]
[0020] Where N represents the total number of pseudo-labels that pass the threshold in the current batch; Indicates the first Category confidence of each pseudo-label; This indicates its target score;
[0021] Average confidence level Represented as:
[0022]
[0023] in, The number of images contained in the current batch. For the first The number of valid pseudo-labels retained in the image. The corresponding confidence score;
[0024] When no predicted bounding boxes in the current batch pass the pseudo-label filtering threshold, the feedback signal for that batch is... .
[0025] Preferably, the specific process of feedback signal smoothing in step S33 includes:
[0026] Feedback metrics Represented as:
[0027]
[0028] Where γ is the smoothing coefficient and t is the current iteration step.
[0029] Preferably, the specific process of step S4, the adaptive dynamic mean teacher update strategy, includes:
[0030] S41. Construct an adaptive EMA decay rate function: The average confidence feedback signal of the pseudo-labels for the current training batch obtained in step S33 is used... During the teacher model parameter update process, an adaptive strategy is used to dynamically calculate the decay rate in the exponential moving average (EMA) update. ;
[0031] S42. Dynamic update of teacher model parameters: using the adaptive EMA decay rate calculated in step S41. The teacher model parameters are updated using an exponential moving average:
[0032]
[0033] in, and These represent the current iteration step. The weights of the teacher and student models at each moment;
[0034] S43. Student Model Optimization and Total Loss Calculation: Minimizing the total loss of the student model on the target domain data. Continuously optimize the student model;
[0035] S44. Linkage Adjustment Mechanism: When the model validation performance meets the preset change conditions, the pseudo-label screening threshold and EMA decay rate corresponding to the current batch are adjusted. Joint adjustments were made; specifically, when model validation performance deteriorated or overfitting trends emerged, the false label screening threshold was increased and / or the EMA decay rate was increased. Enhance the stability of teacher model updates; when model validation performance steadily improves, reduce the false label screening threshold and / or decrease the EMA decay rate. This improves the efficiency of student models in absorbing knowledge from the target domain.
[0036] Preferably, the specific process of constructing the adaptive EMA decay rate function in step S41 includes:
[0037] Adaptive EMA attenuation rate Represented as:
[0038]
[0039] in, The base decay rate has a value that approaches 1. This is the maximum adjustable step size; This is the feedback signal of the average confidence score of the pseudo-labels in the current training batch; In order to align with training rounds Related early dynamic decay control factors.
[0040] Preferably, the specific process of student model optimization and total loss calculation in step S43 includes:
[0041] Total loss The consistency detection loss is composed of pseudo-labels generated by the teacher model:
[0042]
[0043] in, The target domain consistency loss is based on pseudo-labels from the teacher model; The weighting coefficients for consistency loss; Pseudo-labels generated for the teacher model;
[0044] Consistency loss Including classification loss Regression loss :
[0045]
[0046] in, Predict the category for the student model; Predict the bounding box position for the student model; and These represent the pseudo-label category and pseudo-label box position generated by the teacher model, respectively.
[0047] An electronic device includes a memory and a processor, wherein a computer program is stored in the memory, characterized in that: when the processor executes the computer program, it implements any of the steps in the passive domain adaptive target detection method based on pseudo-label confidence feedback described above.
[0048] The beneficial effects of this invention are:
[0049] This invention introduces a pseudo-label confidence feedback mechanism, enabling the model to adaptively adjust during unsupervised self-training and dynamically evaluate and control the quality of pseudo-labels. When the model faces difficult samples or noisy data, it can automatically improve update conservatism, effectively suppressing the adverse effects of low-quality pseudo-labels on model training, thereby significantly reducing the risk of performance degradation caused by noise accumulation.
[0050] This invention proposes an adaptive EMA update strategy based on confidence feedback, overcoming the limitations of traditional fixed momentum update methods. In high-quality pseudo-label scenarios, it can dynamically reduce the decay rate, accelerate the absorption of target domain knowledge by the teacher model, and thus improve cross-domain transfer efficiency and detection accuracy while ensuring model stability.
[0051] This invention designs a dynamic adjustment mechanism that combines the training phase. In the early stage of training, it can quickly absorb the features of the target domain and achieve efficient convergence. In the later stage of training, it can effectively suppress overfitting and avoid the repeated amplification of false label noise, thereby reducing the overall training cost and improving the model's generalization ability.
[0052] Compared with existing technologies, this invention effectively alleviates the "confirmation bias" problem in passive domain adaptive tasks by co-optimizing the pseudo-label screening threshold and the teacher model update parameters. It significantly improves the training stability, noise resistance, and cross-domain target detection performance of the model in the target domain, and has good application value and promotion prospects. Attached Figure Description
[0053] Figure 1 This is a structural diagram of the passive domain adaptive target detection method based on pseudo-label confidence feedback in Embodiment 1 of the present invention;
[0054] Figure 2 This is a flowchart of the adaptive dynamic mean teacher update strategy in Embodiment 1 of the present invention. Detailed Implementation
[0055] Example 1
[0056] The following is a further explanation of the present invention in conjunction with specific embodiments, such as... Figure 1 As shown, this embodiment is a passive domain adaptive target detection method based on pseudo-label confidence feedback, including the following steps:
[0057] S1. Obtain the pre-trained source domain model and unlabeled target domain data: Obtain unlabeled target domain data (such as the Foggy Cityscapes dataset) that contains only image data and no annotation information, and use a model (such as the YOLO model) that has been trained on the source domain (such as the Cityscapes dataset) as the pre-trained source domain model on the target domain.
[0058] S2. Construct the mean teacher self-training architecture: Based on the pre-trained source domain model obtained in step S1, construct the mean teacher self-training architecture, which includes student and teacher models. The student and teacher models have the same network structure, and the initial weights are assigned by the pre-trained source domain model.
[0059] In this embodiment, an SF-YOLO-based backbone network is used as the basic feature extractor and detector head. Specifically, two models with identical network structures are instantiated: the student network and the YOLO backbone network. Teacher Network The weights of the student model are updated via backpropagation using the gradient descent algorithm, while the weights of the teacher model do not participate in backpropagation but are updated only by receiving the weights of the student model and performing an exponential moving average (EMA).
[0060] Among them, student network Includes learnable parameters During self-training, the loss is calculated using pseudo-labels in the target domain, and gradient updates are performed using backpropagation algorithms (such as SGD or AdamW optimizers); teacher network Includes parameters It does not participate in the construction of the computation graph during training, i.e., the gradient is frozen (set as a gradient truncation operation in deep learning frameworks), and its parameters are only passively updated by receiving parameters from the student network.
[0061] Before self-training begins (i.e., during the initialization phase of training), the pre-trained source domain model parameters are loaded completely and equivalently into the system. and In this process, it is ensured that both have the same initial source domain prior knowledge distribution.
[0062] S3. Generate pseudo-labels and calculate confidence feedback signals: Input the image data of the target domain obtained in step S1 into the teacher model constructed in step S2 to generate target prediction boxes and their corresponding confidence scores; dynamically adjust the pseudo-label selection threshold according to the current training state, and obtain pseudo-labels based on the threshold selection; at the same time, calculate the confidence statistics of pseudo-labels in the current training batch, and use them as a feedback signal for dynamic adjustment. The specific process includes:
[0063] S31. Pseudo-label generation and preliminary screening: In each training iteration (Epoch), for the current batch of unlabeled target domain images, weak data augmentation is first performed and then input into the teacher network. The forward propagation is performed to output the initial prediction box and its corresponding class confidence.
[0064] For the j-th image in the current training batch, select the set of predicted bounding boxes whose confidence scores from the teacher model output are greater than the pseudo-label confidence threshold τ. Determine the pseudo-label selection threshold for the current iteration based on the current training state. The initial predicted bounding boxes output by the teacher model are then subjected to non-maximum suppression (NMS) to eliminate overlapping redundant boxes. Further filtering then retains only those with a confidence level greater than the pseudo-label filtering threshold. The predicted bounding boxes are used as a set of high-confidence pseudo-labels. .
[0065] S32. Extract the confidence feedback signal: Calculate the average confidence score of all selected pseudo-labels in the current batch, as a feedback signal reflecting the current performance of the teacher model. And calculate the average confidence level of all valid pseudo-labels in the current batch. This is used as a confidence index to measure the current model's adaptability to the target domain data and the quality of pseudo-labels; the specific process includes:
[0066] Feedback signal Represented as:
[0067]
[0068] Where N represents the total number of pseudo-labels that pass the threshold in the current batch; Indicates the first Category confidence of each pseudo-label; This indicates the target score.
[0069] Average confidence level Represented as:
[0070]
[0071] in, The number of images contained in the current batch. For the first The number of valid pseudo-labels retained in the image. The corresponding confidence score, its value range is theoretically located within... Within the range.
[0072] When no predicted bounding boxes in the current batch pass the pseudo-label filtering threshold, the feedback signal for that batch is... Feedback signal The higher the value, the more reliable the pseudo-labels output by the teacher model are, and the less noise they contain.
[0073] S33. Feedback signal smoothing: The feedback signal is smoothed using a historical moving average to obtain a stable feedback index. :
[0074]
[0075] Where γ is the smoothing coefficient and t is the current iteration step.
[0076] S4. Adaptive Dynamic Mean Teacher Update Strategy: (e.g.) Figure 2 As shown, the student model constructed in step S2 is used to train the target domain data obtained in step S1; in the weight update stage, based on the confidence feedback signal obtained in step S3, the momentum decay rate parameter of the exponential moving average (EMA) is dynamically calculated and adjusted, and the pseudo-label screening threshold and momentum decay rate parameter are adjusted when the model validation performance changes; using the updated momentum decay rate, the weights of the student model are updated to the teacher model. The specific process includes:
[0077] S41. Construct an adaptive EMA decay rate function: The average confidence feedback signal of the pseudo-labels for the current training batch obtained in step S33 is used... During the teacher model parameter update process, an adaptive strategy is used to dynamically calculate the decay rate in the exponential moving average (EMA) update. ;
[0078]
[0079] in, The base decay rate has a value that approaches 1. This is the maximum adjustable step size; This is the feedback signal of the average confidence score of the pseudo-labels in the current training batch; In order to align with training rounds Related early dynamic decay control factors.
[0080] In this embodiment, a novel approach is proposed: extracting the feedback signal through step S3. Dynamic calculation of momentum decay rate The strategy, combined with the control factor during the training phase. The teacher model update rate is adaptively adjusted; and when the model validation performance meets the preset change conditions, the pseudo-label screening threshold and the EMA momentum parameter are jointly adjusted.
[0081] S42. Dynamic update of teacher model parameters: using the adaptive EMA decay rate calculated in step S41. The teacher model parameters are updated using an exponential moving average:
[0082]
[0083] in, and These represent the current iteration step. The weights of the teacher and student models at each moment.
[0084] The aforementioned adaptive computation mechanism enables the model to dynamically adjust the teacher model parameter update rate based on the quality of the current batch of pseudo-labels. When the student model encounters difficult samples that differ significantly from the source domain, the confidence level of the pseudo-labels output by the teacher network is generally low, i.e. Approaching 0, at this point The teacher model is in a highly conservative state to reduce the adverse impact of low-quality pseudo-labels on teacher model updates; conversely, when the model encounters target domain samples that are easier to align with, the pseudo-label confidence is higher. Increase The corresponding reduction, thereby giving Greater weighting prompts teacher networks to more quickly absorb new features extracted from the target domain by student networks.
[0085] S43. Student Model Optimization and Total Loss Calculation: Minimizing the total loss of the student model on the target domain data. We will continue to optimize the student model.
[0086] Total loss The consistency detection loss is composed of pseudo-labels generated by the teacher model:
[0087]
[0088] in, The target domain consistency loss is based on pseudo-labels from the teacher model; The weighting coefficients for consistency loss; Pseudo-labels generated for the teacher model;
[0089] Consistency loss Including classification loss Regression loss :
[0090]
[0091] in, Predict the category for the student model; Predict the bounding box position for the student model; and These represent the pseudo-label category and pseudo-label box position generated by the teacher model, respectively.
[0092] S44. Linkage Adjustment Mechanism: When the model validation performance meets the preset change conditions, the pseudo-label screening threshold and EMA decay rate corresponding to the current batch are adjusted. Joint adjustments were made; specifically, when model validation performance deteriorated or overfitting trends emerged, the false label screening threshold was increased and / or the EMA decay rate was increased. Enhance the stability of teacher model updates; when model validation performance steadily improves, reduce the false label screening threshold and / or decrease the EMA decay rate. This improves the efficiency of student models in absorbing knowledge from the target domain.
[0093] Example 2
[0094] This embodiment sets up an experiment based on the passive domain adaptive target detection method based on pseudo-label confidence feedback proposed in Embodiment 1, and tests and compares it with other existing common mainstream algorithms, and performs quantitative and qualitative analysis of the results.
[0095] In this embodiment, two sets of experiments were set up: one set of experiments with the same scene but different weather (source domain: CityScape dataset, target domain: Foggy CityScape dataset) and another set of experiments with the same weather but different shooting angles (source domain: KITTI dataset, target domain: CityScape dataset).
[0096] The specific parameter settings for the experiment include:
[0097] Maintain conservative base state: setting Δα = 0.009. The baseline value of the EMA momentum parameter for the teacher model is set to 0.999, with a lower bound of 0.990, corresponding to a maximum adjustable step size of 0.009. When the average confidence level of the smoothed batch pseudo-labels is high, the EMA momentum parameter of the teacher model can adaptively decrease from 0.999 to close to 0.990, thereby improving the rate of new knowledge absorption.
[0098] The experimental results are shown in the table below:
[0099] Table 1. Comparison of mAP metrics between the CityScape dataset (source domain) and the Foggy CityScape dataset (target domain).
[0100]
[0101] Table 2 shows the comparison results of mAP metrics for datasets with the KITTI dataset as the source domain and the CityScape dataset as the target domain.
[0102]
[0103] In the two tables above, Source represents the target detection model trained only on the source domain without any domain adaptation processing; DA-FasteR-CNN, EPM, and SF-YOLO are other common mainstream algorithms; Our is the algorithm disclosed in Example 1; Person, Rider, etc. in Table 1 represent the target detection types of each algorithm in the experiment, and the data show their detection accuracy for this type of target; mAP is the average accuracy of each algorithm; as shown in the results in the two tables above, the method in Example 1 of this invention achieved the best results on all indicators, indicating that it has a significant advantage over commonly used mainstream methods in scenarios with large domain offsets.
[0104] The above description is merely a further explanation of the present invention in conjunction with specific embodiments. All descriptions made do not imply any limitation on the scope of protection of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A passive domain adaptive target detection method based on pseudo-label confidence feedback, characterized in that, Includes the following steps: S1. Obtain the pre-trained source domain model and unlabeled target domain data: Obtain unlabeled target domain data that contains only image data and no labeling information, and use the model that has been trained on the source domain as the pre-trained source domain model on the target domain. S2. Construct the mean teacher self-training architecture: Based on the pre-trained source domain model obtained in step S1, construct the mean teacher self-training architecture, which includes a student model and a teacher model. The student model and the teacher model have the same network structure, and the initial weights are assigned by the pre-trained source domain model. S3. Generate pseudo-labels and calculate confidence feedback signals: Input the image data of the target domain obtained in step S1 into the teacher model constructed in step S2 to generate target prediction boxes and their corresponding confidence scores; dynamically adjust the pseudo-label screening threshold according to the current training state, and obtain pseudo-labels based on the threshold; at the same time, calculate the confidence statistics of pseudo-labels in the current training batch, and use them as feedback signals for dynamic adjustment. S4. Adaptive dynamic mean teacher update strategy: Use the student model constructed in step S2 to train the target domain data obtained in step S1. During the weight update phase, based on the confidence feedback signal obtained in step S3, the momentum decay rate parameter of the exponential moving average (EMA) is dynamically calculated and adjusted, and the pseudo-label screening threshold and the momentum decay rate parameter are adjusted when the model validation performance changes; the updated momentum decay rate is used to update the weights of the student model to the teacher model.
2. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 1, characterized in that: The network structure in step S2, which constructs the mean teacher self-training architecture, is based on the SF-YOLO object detection model. The weights of the student model are updated through backpropagation using the gradient descent algorithm, while the weights of the teacher model do not participate in backpropagation but are updated by receiving the weights of the student model and performing an exponential moving average (EMA) update.
3. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 2, characterized in that: The specific process of generating pseudo-labels and calculating confidence feedback signals in step S3 includes: S31. Pseudo-label generation and preliminary screening: Input the target domain image into the teacher model to obtain the predicted bounding box, class probability, and target score; for the j-th image in the current training batch, select the set of predicted bounding boxes whose confidence level output by the teacher model is greater than the pseudo-label confidence screening threshold τ. Determine the pseudo-label filtering threshold for the current batch based on the current training status. The prediction results are then filtered using the pseudo-label filtering threshold, retaining high-confidence prediction boxes as pseudo-labels for the current batch. ; S32. Extract the confidence feedback signal: Calculate the average confidence score of all selected pseudo-labels in the current batch, as a feedback signal reflecting the current performance of the teacher model. And calculate the average confidence level of all valid pseudo-labels in the current batch. This is used as a confidence index to measure the current model's adaptability to the target domain data and the quality of the pseudo-labels; S33. Feedback signal smoothing: The feedback signal is smoothed using a historical moving average to obtain a stable feedback index. .
4. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 3, characterized in that: The specific process of extracting the confidence feedback signal in step S32 includes: The feedback signal Represented as: Where N represents the total number of pseudo-labels that pass the threshold in the current batch; Indicates the first Category confidence of each pseudo-label; This indicates its target score; The average confidence level Represented as: in, The number of images contained in the current batch. For the first The number of valid pseudo-labels retained in the image. The corresponding confidence score; When no predicted bounding box in the current batch passes the pseudo-label filtering threshold, the feedback signal for that batch is... .
5. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 4, characterized in that: The specific process of the feedback signal smoothing process in step S33 includes: The feedback metrics Represented as: Where γ is the smoothing coefficient and t is the current iteration step.
6. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 5, characterized in that: The specific process of step S4, the adaptive dynamic mean teacher update strategy, includes: S41. Construct an adaptive EMA decay rate function: The average confidence feedback signal of the pseudo-labels for the current training batch obtained in step S33 is used... During the teacher model parameter update process, an adaptive strategy is used to dynamically calculate the decay rate in the exponential moving average (EMA) update. ; S42. Dynamic update of teacher model parameters: using the adaptive EMA decay rate calculated in step S41. The teacher model parameters are updated using an exponential moving average: in, and These represent the current iteration step. The weights of the teacher and student models at each moment; S43. Student Model Optimization and Total Loss Calculation: Minimizing the total loss of the student model on the target domain data. Continuously optimize the student model; S44. Linkage Adjustment Mechanism: When the model validation performance meets the preset change conditions, the pseudo-label screening threshold and EMA decay rate corresponding to the current batch are adjusted. Joint adjustments were made; specifically, when model validation performance deteriorated or overfitting trends emerged, the false label screening threshold was increased and / or the EMA decay rate was increased. Enhance the stability of teacher model updates; when model validation performance steadily improves, reduce the false label screening threshold and / or decrease the EMA decay rate. This improves the efficiency of student models in absorbing knowledge from the target domain.
7. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 6, characterized in that: The specific process of constructing the adaptive EMA decay rate function in step S41 includes: The adaptive EMA attenuation rate Represented as: in, Based on the attenuation rate, This is the maximum adjustable step size; This is the feedback signal of the average confidence score of the pseudo-labels in the current training batch; In order to align with training rounds Related early dynamic decay control factors.
8. The passive domain adaptive target detection method based on pseudo-label confidence feedback according to claim 7, characterized in that: The specific process of student model optimization and total loss calculation in step S43 includes: The total loss The consistency detection loss is composed of pseudo-labels generated by the teacher model: in, The target domain consistency loss is based on pseudo-labels from the teacher model; The weighting coefficients for consistency loss; Pseudo-labels generated for the teacher model; The consistency loss Including classification loss Regression loss : in; Predict the category for the student model; Predict the bounding box position for the student model; and These represent the pseudo-label category and pseudo-label box position generated by the teacher model, respectively.
9. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements any step in the passive domain adaptive target detection method based on pseudo-label confidence feedback as described in any one of claims 1-8.