Multi-expert progressive distribution transfer method for passive target detection

Abstract

The application discloses a multi-expert gradual distribution migration method for passive target detection and belongs to the technical field of computer vision target detection. The method comprises the following steps: decoupling a source domain pre-trained basic detector into a shared feature extractor and a plurality of parallel expert units with logical order, and constructing a student detector and a teacher detector according to the same; predicting pseudo labels of unlabeled target domain images by using the teacher detector, and dividing the pseudo labels into four gradual levels, i.e., a reliable layer, a potential layer, an unknown layer and a background layer according to confidence; calculating a comprehensive consistency matrix of a current order expert unit and a previous order expert unit, tracking the pseudo labels across expert sources, eliminating noise samples that do not meet the consistency and performing cross-expert fusion; adaptively and self-training the student detector based on the corrected and fused pseudo labels and iteratively optimizing the student detector, and dynamically updating the teacher detector based on an exponential moving average mechanism. The application can effectively suppress the accumulation of pseudo label errors and significantly improve the cross-domain detection precision.

Description

Technical Field

[0001] This invention belongs to the field of computer vision target detection technology, specifically involving a multi-expert progressive distributed migration method for passive target detection. Background Technology

[0002] With the continuous development of technology, object detection technology has become relatively mature. However, a core pain point in practical applications is the limited access to source domain data, which has given rise to the field of passive object detection (SFOD). Its core requirement is to achieve cross-domain adaptation using only pre-trained models in the source domain and unlabeled target domain data, without any access to the source domain labeled data. This requirement is particularly urgent in scenarios with strict data privacy protection, high annotation costs, and limited transmission.

[0003] Currently, most mainstream SFOD technologies are based on the teacher-student model paradigm, achieving model adaptation through knowledge distillation, feature alignment, or pseudo-label refinement. However, these technologies have many shortcomings that severely limit detection performance and applicability. Specific defects are as follows: (1) Differences in data distribution between the source and target domains. Significant differences in data distribution between the source and target domains cause the detection model to make prediction errors in the target domain when predicting false labels, or to result in bounding box bias. These factors severely affect the effectiveness of existing methods.

[0004] (2) Noise accumulation effect. Existing passive domain adaptive methods inevitably introduce certain noise during pseudo-label prediction. As training iterations continue, the noise will gradually accumulate and amplify, thereby interfering with model optimization and even causing a decrease in detection performance.

[0005] (3) Unified modeling of latent knowledge mining and noise accumulation effect in the target domain. In passive domain adaptive tasks, there is a common distribution difference between the source domain and the target domain, accompanied by the gradual accumulation of noise introduced by pseudo-labels. However, existing methods usually only focus on the optimization of a single problem, lacking in-depth modeling and unified analysis of the inherent coupling mechanism between the two. This results in a trade-off between latent knowledge mining and noise suppression in the target domain, making it difficult to effectively take both into account at the same time, thus limiting the overall performance improvement.

[0006] Existing SFOD technology still has key shortcomings in modeling the differences in the distribution between the source and target domains, suppressing the continuous accumulation of pseudo-label noise, and unifying the modeling of potential knowledge mining and noise propagation processes in the target domain, making it difficult to meet the needs of high-precision target detection in unlabeled scenarios.

[0007] Therefore, there is an urgent need to propose new technical solutions to overcome the aforementioned bottlenecks. Based on this, this invention proposes a multi-expert progressive distributed migration method for passive target detection. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a multi-expert progressive distributed migration method for passive target detection, thereby solving the problems in existing technologies.

[0009] The objective of this invention can be achieved through the following technical solutions: A multi-expert progressive distributed transfer method for passive target detection includes the following steps: S1, decouple the source domain pre-trained static base detector into a shared feature extractor and an expert group module composed of multiple parallel expert units, wherein each expert unit consists of an independent private feature extractor and a detection head, and configure a logical order for the multiple parallel expert units; construct a student detector and a teacher detector based on each decoupled module, and initialize them using the weights of the static base detector; S2, using the teacher detector to perform target detection prediction on the unlabeled target domain training image, and obtain the predicted pseudo-label of each expert unit in the teacher network; according to the confidence threshold level, the predicted pseudo-label of each expert unit is divided into four progressive layers, including reliable layer, potential layer, unknown layer and background layer, to characterize the reliability evolution state of the sample. S3, for the current expert unit with a priority of k, calculate the category-level consistency matrix and the bounding box-level consistency matrix between it and the preceding expert unit with a priority of k-1, and combine the two to generate a comprehensive consistency matrix; based on the comprehensive consistency matrix, perform cross-expert source tracking on the pseudo-labels in the reliable layer and the potential layer of the current expert unit; remove noise samples that do not meet the consistency source tracking, obtain the corrected pseudo-labels, and perform cross-expert fusion on the corrected pseudo-labels based on the historical results of the preceding expert units; S4. Based on the corrected and fused pseudo-labels, the expert unit and the shared feature extractor in the student detector are self-trained and iteratively optimized using the object detection loss function; at the same time, the parameters of the teacher detector are dynamically updated based on the optimized parameters of the student detector using an exponential moving average mechanism.

[0010] Further, in S1, the shared feature extractor and the private feature extractor are constructed based on the VGG backbone network, wherein the shared feature extractor corresponds to the network layer located before layer 19 in the VGG backbone network, and the private feature extractor of each expert unit corresponds to layers 19 to 20 of the VGG backbone network; the detection head is composed of the Region Candidate Network (RPN) and the RCNN module in the Faster R-CNN detector.

[0011] Furthermore, in S2, the specific rules for dividing the four progressive levels are as follows: When predicting the confidence probability of a sample When, it is divided into the aforementioned reliable layer; When predicting the confidence probability of a sample When, it is divided into the potential layer; When predicting the confidence probability of a sample When, it is divided into the unknown layer; When predicting the confidence probability of a sample At that time, it is divided into the background layer; in, and It is a preset threshold constant, and .

[0012] Furthermore, in S3, the category-level consistency matrix between the current expert unit with priority k and the preceding expert unit with priority k-1 is calculated. The formula is: in, This represents the category result of the i-th predicted target in the reliable layer of the expert unit with priority k. This represents the category result of the j-th predicted target in the latent layer of the expert unit with a priority of k-1. A value greater than 0 indicates that the corresponding target category belongs to the foreground.

[0013] Furthermore, in S3, the bounding box-level consensus matrix between the current expert unit with priority k and the preceding expert unit with priority k-1 is calculated. The formula is: in, Represents the intersection-union ratio function. This represents the bounding box of the i-th predicted target in the reliable layer of the expert unit with priority k. This represents the bounding box of the j-th predicted target in the latent layer of the expert unit with a priority of k-1. This represents the preset overlap threshold constant.

[0014] Furthermore, in S3, the process of cross-expert source tracking and screening based on the comprehensive consistency matrix includes: The category-level consistency matrix With the bounding box level consistency matrix Multiplication yields the comprehensive consistency matrix. ; when When the i-th reliable layer sample of the expert unit with a ranking of k can be derived from the j-th potential layer sample of the expert unit with a ranking of k-1, it is retained; otherwise, it is discarded as a noise sample. Using the same screening strategy, the compatibility matrix between the latent layer of the expert unit with rank k and the unknown layer of the expert unit with rank k-1 is calculated. The pseudo-labels in the latent layer are corrected for consistency constraints, and latent layer samples that do not meet the consistency constraints are removed.

[0015] Furthermore, in S3, the cross-expert fusion process includes: For the current expert unit with a priority of k, its corresponding reliable layer sample is obtained by performing non-maximum suppression (NMS) fusion based on the corrected reliable layer samples output by all k-1 expert units preceding it; its corresponding potential layer sample is obtained by directly inheriting the corrected potential prediction results from all k-1 expert units preceding it. Specifically, for the first expert unit with a priority of 0, its reliable layer and potential layer samples are directly derived from the prediction results of the static base detector; for the expert unit with a priority of 1, its reliable layer and potential layer are respectively aggregated from the correction samples of the corresponding levels in the static base detector and the expert unit with a priority of 0.

[0016] Furthermore, in S4, the self-training iterative optimization process also includes: the shared feature extractor simultaneously receives joint gradient feedback from the backpropagation of all parallel expert units in the student detector, and realizes collaborative learning and feature optimization of the shared feature extractor based on the joint gradient feedback.

[0017] A computer storage medium storing a readable program that, when executed, instructs a computing device to perform the multi-expert progressive distributed migration method for passive target detection as described above.

[0018] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform operations corresponding to the multi-expert progressive distributed migration method for passive target detection described above.

[0019] The beneficial effects of this invention are: 1. This invention uses the idea of ​​gradual transition of sample distribution to track and predict the evolution of pseudo-label distribution, thereby effectively ensuring that high-quality and stable samples dominate the training process. At the same time, it gradually introduces potential samples with target domain style, thereby achieving robust modeling and stable fusion of target domain knowledge.

[0020] 2. This invention systematically divides the overall predicted sample distribution and further performs fine-grained modeling of robust samples, potential target domain knowledge samples, and noise and background samples, thereby more accurately depicting the evolution of the predicted sample distribution during passive domain adaptive training and realizing the effective mining and learning of potential knowledge in the target domain.

[0021] 3. To address the shortcomings of existing technologies, such as the difference in data distribution between the source and target domains and the accumulation of errors, this invention constructs a cross-expert transfer consistency module based on the idea of ​​multi-expert progressive distributed transfer. This module automatically removes and corrects erroneous pseudo-labels, effectively reducing misclassification and missed detections, and providing highly reliable supervision for model optimization, thus fundamentally solving the pain points of the self-training paradigm.

[0022] 4. This invention proposes a passive domain adaptive strategy based on multi-expert progressive distributed transfer, which effectively improves the detector's adaptive ability to the target domain distribution while significantly enhancing the stability and robustness of the training process.

[0023] 5. This invention abandons the limitations of existing SFOD models that rely excessively on teacher-student models, and innovatively proposes the idea of ​​"multi-expert collaboration + gradual distributed migration" to build a complete and reusable technology system. It promotes the transformation of SFOD technology from source domain dependent to target domain adapted, provides a new research direction for the field, and drives the technological upgrading of related industries. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the Multi-Expert Progressive Distributed Transfer Framework (ME-PDT) for Passive Target Detection of the present invention; Figure 2 This is a schematic diagram illustrating the core idea of ​​the progressive migration of sample distribution proposed in this invention; Figure 3 The graph shows the performance curves of the detector during the training process of this invention in two domain adaptive scenarios (C→F and M→C). Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 Multi-expert progressive distributed transfer framework (ME-PDT) for passive target detection, such as Figure 1 As shown, WA represents Weak Augmentation, SA represents Strong Augmentation, and CETC represents the Cross-Expert Transition Consistency module. The ME-PDT framework uses "multi-expert collaboration + progressive distributed transfer" as its core logic. It does not rely on source domain labeled data throughout the process, but only on pre-trained source domain models and unlabeled target domain data. Through modular design and collaborative optimization, it achieves a performance breakthrough in passive target detection. Its core technologies focus on the engineering construction of the model architecture, the technical implementation of each expert module, and the implementation of the progressive optimization mechanism.

[0028] A multi-expert progressive distributed transfer learning method for passive object detection utilizes unlabeled target domain data to perform cross-domain adaptive training on a source domain pre-trained model, including the following steps: S1, decouple the basic detector and build a multi-expert student and teacher detector; exist Figure 1 In this study, the Faster R-CNN detector is used as the base detector for ease of theoretical explanation and evaluation. First, the base detector, already trained in the source domain, is decoupled into a shared feature extractor and an expert group module consisting of three parallel expert units. Each expert unit comprises an independent private feature extractor and a detection head. Furthermore, the private feature extractor, based on layers 19 and 20 of the VGG backbone network, is used to learn high-level semantic representations specific to each expert. The detection head consists of the Region Proposal Network (RPN) and R-CNN module from the Faster R-CNN detector, used to generate candidate regions, classify objects, and regress bounding boxes, thereby achieving differentiated modeling and refined detection of object distributions by different experts.

[0029] Subsequently, student and teacher detectors are constructed based on the decoupled detector modules, and initialized using the weights of the base detectors pre-trained in the source domain to fully inherit source domain knowledge and accelerate model convergence. This initialization strategy also provides a stable starting point for subsequent adaptive training in the source domain, thereby improving overall training efficiency and performance ceiling.

[0030] S2, obtain the predicted pseudo-label distribution and perform hierarchical partitioning; Based on the static base detector and the teacher detector, non-maximum suppression (NMS) is used to predict the target domain training images, with the confidence threshold set to 0.2 and the bounding box overlap ratio (IoU) threshold set to 0.5. This yields the pseudo-label prediction results of the static teacher network and each expert module of the teacher network, denoted as follows: , .in Let represent the prediction structure of the j-th expert unit, and let ci, bi, and pi represent the category label, bounding box, and confidence probability of the i-th prediction result, respectively.

[0031] Subsequently, the pseudo-label distribution hierarchy is divided using the following formula, primarily classifying the predicted pseudo-labels into four levels: a reliable layer, a latent layer, an unknown layer, and a background layer. The reliable layer mainly contains robust and reliable predicted samples, the latent layer contains rich target domain semantics, the unknown layer consists of a large amount of foreground semantics and some effective target domain semantics, and the background layer mainly consists of background elements. Based on this, the evolution of sample reliability from low to high can be further characterized during the training process.

[0032] in , , is a threshold constant used to delineate the reliable layer, potential layer, unknown layer, and background layer regions. p is the confidence probability of the predicted sample, and L represents the distribution layer number identifier.

[0033] S3, construct a cross-expert transfer consistency module to realize cross-expert prediction pseudo-label distribution state tracking, and then complete multi-expert prediction pseudo-label correction and fusion; In the Cross-Expert Transition Consistency (CETC) module, the cross-expert transition consistency formula includes: a category-level consistency matrix and a box-level consistency matrix; The formula for the category-level consistency matrix is: The formula for the frame-level consistency matrix is: The formula for the overall consistency matrix is: in, This represents the intersection-union ratio of two bounding boxes, indicating the degree of overlap between the predicted result and the actual annotation. This represents the category of the i-th predicted target in the k-th expert reliability layer. This represents the category of the (j-1)th predicted target in the (k-1)th expert potential layer. It is a threshold constant used to measure the overlap of rectangles. This indicates that the i-th reliable layer sample of the k-th expert and the j-th potential layer sample of the (k-1)-th expert correspond to the same target instance. Based on this condition, a source-tracking matrix is ​​constructed for the reliable layer samples of the k-th expert. This is used to determine whether each reliable sample can be traced back to the latent layer prediction of the (k-1)th expert. If a sample can establish a consistent correspondence, it is considered stable and reliable, and is thus retained as a corrected reliable layer sample; otherwise, it is considered to originate from a low-confidence region and has a higher risk of noise, and is therefore removed from the corrected reliable set.

[0034] Similarly, we apply the same asymptotic correction process to the latent layer samples of the k-th expert. Specifically, we compute the compatibility matrix between the latent layer of the k-th expert and the uncertain layer of the (k-1)-th expert. This is used to determine whether each potential sample has a stable source from the previous layer of expert uncertainty. Samples that satisfy the consistency constraint are retained as corrected potential layer samples; otherwise, they are discarded.

[0035] Following cross-expert correction, the corrected reliable layer samples and latent layer samples are further fused across experts to achieve progressive inheritance and expansion of stable knowledge. Specifically, the reliable layer samples of the k-th expert are obtained by fusing the corrected reliable prediction results of the first k-1 experts using non-maximum suppression (NMS), thereby integrating consistent and stable high-confidence target information across different stages. Simultaneously, the latent layer samples inherit from the corrected latent prediction results of the first k-1 experts, used to retain candidate targets containing semantic knowledge of the target domain. The fused reliable and latent samples, along with their corresponding pseudo-labels, are used together to train the k-th expert module, enabling the detector to gradually absorb stable knowledge during iteration and continuously expand towards the target domain distribution. It should be noted that for the 0th expert, both its reliable and latent layer samples come from the prediction results of the pre-trained base detector in the static source domain; while for the 1st expert, its reliable and latent layers are respectively aggregated from the correction samples of the base detector and the corresponding level of the 0th expert.

[0036] S4, the passive domain adaptive training phase, uses the corrected and fused pseudo-labels to iteratively optimize the student detector, while the exponential moving average (EMA) mechanism is used to update the teacher detector to stabilize the training process and improve the quality of pseudo-labels. Given a target domain training image xt, the predicted pseudo-label after cross-expert transfer consistency correction and fusion by the k-th expert is: First, the k-th expert module and the shared extractor in the student detector are optimized using the built-in detection loss function (denoted as Ldet) of the Faster-RCNN detector.

[0037] in For the parameters of the k-th expert module of the student detector, Extractor parameters are shared for student detectors. Most importantly, an exponential moving average (EMA) mechanism is used to update the teacher detector.

[0038] Example 2 In this embodiment, the performance of the Multi-Expert Progressive Distributed Migration Framework (ME-PDT) proposed in Example 1 is verified and analyzed. To verify the practical technical effectiveness of the Multi-Expert Progressive Distributed Transfer Model (ME-PDT) framework proposed in this invention, this embodiment constructs three sets of cross-domain transfer test scenarios while fully adhering to the core assumptions of SFOD (i.e., without accessing source domain labeled data). All three experiments use Faster R-CNN (FRCNN) as the base detector, and a "SourceOnly" model trained solely on source domain data and tested directly on the target domain is used as the performance lower bound benchmark.

[0039] Experiment 1: Migration Test in Complex Weather Environments (C→F Adaptation Experiment) This experiment simulates the distribution migration from normal weather conditions to complex foggy conditions. Using a teacher model pre-trained under normal weather conditions, the ME-PDT framework of this invention is applied to perform multi-expert collaborative training and progressive distribution migration on unlabeled target domain data in foggy conditions. During the testing phase, the model's mean accuracy (AP) and overall mean (mAP) are evaluated for eight fine-grained target categories (person, rider, car, truck, bus, train, mbike, bicycle).

[0040] The experimental results are shown in Package 1: Table 1: Experimental results of AP (%) in the C→F adaptation experiment.

[0041] As shown in Table 1, the SourceOnly method without transfer is severely affected by the bias of source domain weather features, with an mAP of only 18.2%. While the performance is improved after introducing existing mainstream passive target detection algorithms, bottlenecks still exist for some categories. Using the ME-PDT framework of this invention, not only does it perform excellently in single categories (e.g., AP reaches 54.7% for cars and 44.3% for bicycles), but the overall mAP also significantly increases to 42.0%, achieving the best global result. This demonstrates that the feature transfer expert module of this invention can effectively decouple cross-domain feature interference caused by weather environment, significantly improve the model's fine-grained target recognition accuracy under complex environmental changes, and break the limitations of source domain knowledge binding.

[0042] Experiment 2: Cross-camera and cross-scene transfer test (K→C experiment) This experiment aims to verify the model's generalization ability across different acquisition devices and urban scenes. The model, pre-trained on a dataset of the first type of urban scene (or camera configuration), is unsupervisedly transferred to a dataset of the second type of dense urban scene. After feature alignment and pseudo-label refinement using the ME-PDT framework, the model's overall MAP score in the target domain is calculated.

[0043] The experimental results are shown in Table 2: Table 2: Percentage of AP usage (%) in K→C experiments As shown in Table 2, the domain differences across devices / scenes resulted in a MAP of only 39.6% for SourceOnly. Most leading SFOD baseline methods (such as PETS and LPU) have MAPs ranging from 47.0% to 48.4%, with the highest leading baseline LPLD at 51.3%. The ME-PDT method proposed in this application overcomes this bottleneck, further improving the target domain detection accuracy to a MAP of 52.0%. This demonstrates that by relying on the dynamic threshold adjustment network of the pseudo-label refinement expert module and feature consistency verification, this scheme can effectively filter out noisy pseudo-labels generated by sensor or viewpoint changes during cross-scene migration, providing high-quality supervision signals and exhibiting excellent cross-device and cross-scene adaptability.

[0044] Experiment 3: Large-span extreme distribution migration test (M→C experiment) This experiment sets a highly challenging domain gap, simulating extreme distribution migration from synthetic data / specific viewpoints (source domain) to real-world images (target domain). Under these significant inter-domain differences, the ME-PDT framework's three-stage progressive optimizer is activated, coupled with the Cross-Expert Transfer Consistency (CETC) module to track and eliminate unstable samples, ultimately outputting target detection results and evaluating the MAP (Map Value).

[0045] Table 3: Percentage of AP usage in the M→C experiment (%) As shown in Table 3, the significant modal differences caused the performance of the SourceOnly model to degrade dramatically to 32.0%. However, under equally challenging transfer conditions, the ME-PDT framework of this invention maintained extremely high robustness, achieving a top MAP score of 52.3%, significantly outperforming suboptimal baseline methods (such as LPLD at 49.4%). This demonstrates that under extreme cross-domain conditions, the feature discrimination expert of this invention ensures high-purity feature representation through the collaborative optimization of feature clustering and decision boundaries. Furthermore, the progressive distributed transfer effectively avoids drastic oscillations in the model during training, proving that this scheme can still achieve accurate and stable object detection even when there is a significant domain gap between the source and target domains.

[0046] Experiment 4: Robustness Test of Passive Domain Adaptive Training Performance This experiment evaluates the effect of the proposed method on performance robustness during the training process based on C→F and K→C domain adaptive scenarios. Figure 3 This paper compares the detector performance changes during training with some existing passive domain adaptive methods and the method of this invention. Figure 3 As can be seen, existing passive domain adaptive methods are prone to performance degradation during training, mainly due to the accumulation of noise in pseudo-labels during training iterations. However, the method proposed in this invention not only significantly improves the detector's performance in the target domain but also effectively ensures the robustness of performance during training.

[0047] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0048] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A multi-expert progressive distributed migration method for passive target detection, characterized in that, Includes the following steps: S1, decouple the source domain pre-trained static base detector into a shared feature extractor and an expert group module composed of multiple parallel expert units, wherein each expert unit consists of an independent private feature extractor and a detection head, and configure a logical order for the multiple parallel expert units; construct a student detector and a teacher detector based on each decoupled module, and initialize them using the weights of the static base detector; S2, using the teacher detector to perform target detection prediction on the unlabeled target domain training image, and obtain the predicted pseudo-label of each expert unit in the teacher network; according to the confidence threshold level, the predicted pseudo-label of each expert unit is divided into four progressive layers, including reliable layer, potential layer, unknown layer and background layer, to characterize the reliability evolution state of the sample. S3, for the current expert unit with the order of k, calculate the category-level consistency matrix and the bounding box-level consistency matrix between it and the preceding expert unit with the order of k-1, and combine the two to generate a comprehensive consistency matrix; Based on the comprehensive consistency matrix, cross-expert source tracking is performed on the pseudo-labels in the reliable layer and the potential layer of the current expert unit; noisy samples that do not meet the consistency source tracking are removed to obtain the corrected pseudo-labels, and cross-expert fusion is performed on the corrected pseudo-labels based on the historical results of the previous expert units. S4. Based on the corrected and fused pseudo-labels, the expert unit and the shared feature extractor in the student detector are self-trained and iteratively optimized using the object detection loss function; at the same time, the parameters of the teacher detector are dynamically updated based on the optimized parameters of the student detector using an exponential moving average mechanism.

2. The multi-expert progressive distributed migration method for passive target detection according to claim 1, characterized in that, In S1, the shared feature extractor and the private feature extractor are constructed based on the VGG backbone network, wherein the shared feature extractor corresponds to the network layer in the VGG backbone network before layer 19, and the private feature extractor of each expert unit corresponds to layers 19 to 20 of the VGG backbone network; the detection head is composed of the Region Candidate Network (RPN) and the RCNN module in the Faster R-CNN detector.

3. The multi-expert progressive distributed migration method for passive target detection according to claim 1, characterized in that, In S2, the specific rules for dividing the four progressive levels are as follows: When predicting the confidence probability of a sample When, it is divided into the aforementioned reliable layer; When predicting the confidence probability of a sample When, it is divided into the potential layer; When predicting the confidence probability of a sample When, it is divided into the unknown layer; When predicting the confidence probability of a sample At that time, it is divided into the background layer; in, and It is a preset threshold constant, and .

4. The multi-expert progressive distributed migration method for passive target detection according to claim 3, characterized in that, In S3, the category-level consistency matrix between the current expert unit with a priority of k and the preceding expert unit with a priority of k-1 is calculated. The formula is: in, This represents the category result of the i-th predicted target in the reliable layer of the expert unit with priority k. This represents the category result of the j-th predicted target in the latent layer of the expert unit with a priority of k-1. A value greater than 0 indicates that the corresponding target category belongs to the foreground.

5. The multi-expert progressive distributed migration method for passive target detection according to claim 4, characterized in that, In S3, the bounding box-level consensus matrix between the current expert unit with priority k and the preceding expert unit with priority k-1 is calculated. The formula is: in, Represents the intersection-union ratio function. This represents the bounding box of the i-th predicted target in the reliable layer of the expert unit with priority k. This represents the bounding box of the j-th predicted target in the latent layer of the expert unit with a priority of k-1. This represents the preset overlap threshold constant.

6. The multi-expert progressive distributed migration method for passive target detection according to claim 5, characterized in that, In S3, the process of cross-expert source tracking and screening based on the comprehensive consistency matrix includes: The category-level consistency matrix With the bounding box level consistency matrix Multiplication yields the comprehensive consistency matrix. ; when When the i-th reliable layer sample of the expert unit with a ranking of k can be derived from the j-th potential layer sample of the expert unit with a ranking of k-1, it is retained; otherwise, it is discarded as a noise sample. Using the same screening strategy, the compatibility matrix between the latent layer of the expert unit with rank k and the unknown layer of the expert unit with rank k-1 is calculated. The pseudo-labels in the latent layer are corrected for consistency constraints, and latent layer samples that do not meet the consistency constraints are removed.

7. The multi-expert progressive distributed migration method for passive target detection according to claim 1, characterized in that, In S3, the cross-expert fusion process includes: For the current expert unit with a priority of k, its corresponding reliable layer sample is obtained by performing non-maximum suppression (NMS) fusion based on the corrected reliable layer samples output by all k-1 expert units preceding it; its corresponding potential layer sample is obtained by directly inheriting the corrected potential prediction results from all k-1 expert units preceding it. Specifically, for the first expert unit with a priority of 0, its reliable layer and potential layer samples are directly derived from the prediction results of the static base detector; for the expert unit with a priority of 1, its reliable layer and potential layer are respectively aggregated from the correction samples of the corresponding levels in the static base detector and the expert unit with a priority of 0.

8. The multi-expert progressive distributed migration method for passive target detection according to claim 1, characterized in that, In S4, the self-training iterative optimization process further includes: the shared feature extractor simultaneously receives joint gradient feedback from the backpropagation of all parallel expert units in the student detector, and realizes collaborative learning and feature optimization of the shared feature extractor based on the joint gradient feedback.

9. A computer storage medium storing a readable program, characterized in that, When the program runs, it can instruct the computing device to perform the multi-expert progressive distributed migration method for passive target detection as described in any one of claims 1-8.

10. An electronic device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the multi-expert progressive distributed migration method for passive target detection as described in any one of claims 1-8.

Images