Adaptive target detection method and device, electronic equipment, storage medium and program product

Abstract

The application provides a kind of adaptive target detection method, device, electronic equipment, storage medium and program product, it is related to target detection technical field, the method comprises: based on source domain detection model, initialize student model and teacher model;Based on original target domain image, determine the first prediction probability distribution of candidate target frame by the teacher model, and determine the second prediction probability distribution of the candidate target frame by the student model;Image block corresponding to the candidate target frame is input into visual semantic teacher model, and third prediction probability distribution is obtained;Based on the first prediction probability distribution and the third prediction probability distribution, determine pseudo label;Based on the pseudo label and the second prediction probability distribution, the student model is updated, and cross-domain target detection model is obtained, to realize adaptive target detection based on the cross-domain target detection model.The model precision of cross-domain target detection model can be effectively improved by the application.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to an adaptive target detection method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] In recent years, deep learning models have been widely deployed in real-world online scenarios such as autonomous driving, security monitoring, robotics, and remote sensing, significantly improving the processing efficiency and accuracy of various visual tasks. However, during actual deployment, models often encounter data distribution shifts due to factors such as weather changes, differences in imaging equipment, and sudden changes in scene style, leading to a significant drop in the performance of pre-trained models on target domain data. To address this issue, Test-Time Adaptation (TTA) technology has emerged. Its core objective is to allow pre-trained models to adapt to new data distributions in real time during the testing phase, which relies solely on unlabeled target domain data streams. However, this traditional adaptive technique can only simulate supervision signals by generating pseudo-labels on the model itself. In complex tasks such as object detection, due to the coupling between classification and localization regression, pseudo-label noise is easily accumulated in the early stages of strong domain shifts, leading to model instability during training and affecting the accuracy of the final object detection model. Summary of the Invention

[0003] This invention provides an adaptive target detection method, device, electronic device, storage medium, and program product to solve the problem that the use of traditional adaptive techniques in the prior art can easily lead to instability in the model training process, which affects the accuracy of the final target detection model.

[0004] This invention provides an adaptive target detection method, comprising the following steps: Based on the source domain detection model, initialize the student model and the teacher model; Based on the original target domain image, the teacher model determines the first predicted probability distribution of the candidate target boxes, and the student model determines the second predicted probability distribution of the candidate target boxes. The image patch corresponding to the candidate target box is input into the visual semantic teacher model to obtain the third prediction probability distribution; Based on the first predicted probability distribution and the third predicted probability distribution, pseudo-labels are determined; Based on the pseudo-labels and the second predicted probability distribution, the student model is updated to obtain a cross-domain target detection model, so as to achieve adaptive target detection based on the cross-domain target detection model.

[0005] According to an adaptive object detection method provided by the present invention, the method comprises determining a first predicted probability distribution of candidate object boxes based on an original object domain image using the teacher model and a second predicted probability distribution of the candidate object boxes using the student model, including: Obtain the original target domain image from the target domain test data; The original target domain image is processed to obtain a weakly enhanced image and a strongly enhanced image; The weakly enhanced image is input into the teacher model to obtain candidate target boxes and the corresponding first prediction probability distribution; The enhanced image is input into the student model to obtain candidate target boxes and corresponding second prediction probability distributions.

[0006] According to an adaptive object detection method provided by the present invention, the step of inputting the image patch corresponding to the candidate object box into a visual semantic teacher model to obtain a third predicted probability distribution includes: Obtain integrated text hints for the category to be detected, wherein the integrated text hints include multiple text hints of different styles; The image patch corresponding to the candidate target box is cropped from the original target domain image; The integrated text and the image patch are input into the visual semantic teacher model to obtain the third prediction probability distribution.

[0007] According to an adaptive object detection method provided by the present invention, the step of inputting the integrated text and the image patch into the visual semantic teacher model to obtain a third predicted probability distribution includes: The integrated text prompt is input into the text encoder of the visual semantic teacher model to generate the category text features of the category to be detected; The image blocks are input into the image encoder of the visual semantic teacher model to generate image features; Calculate the feature similarity between the image features and the category text features of each category to be detected, and determine the third prediction probability distribution based on the feature similarity.

[0008] According to an adaptive target detection method provided by the present invention, determining pseudo-labels based on a first predicted probability distribution and a third predicted probability distribution includes: Based on the first predicted probability distribution and the third predicted probability distribution, a fusion probability distribution is determined; When the confidence level corresponding to the fusion probability distribution is determined to be greater than a preset threshold, a pseudo-label is determined based on the fusion probability distribution.

[0009] According to an adaptive target detection method provided by the present invention, determining a fused probability distribution based on a first predicted probability distribution and a third predicted probability distribution includes: If the highest probability index of the category of the first predicted probability distribution is consistent with that of the third predicted probability distribution, then the first predicted probability distribution is used as the fusion probability distribution. If the highest probability index of the category of the first predicted probability distribution is inconsistent with that of the second predicted probability distribution, then the first predicted probability distribution and the third predicted probability distribution are weighted and fused to obtain a fused probability distribution.

[0010] According to an adaptive object detection method provided by the present invention, the step of updating the student model based on the pseudo-label and the second predicted probability distribution to obtain a cross-domain object detection model includes: Based on the pseudo-label and the second predicted probability distribution, a joint loss is determined, wherein the joint loss includes a detection loss and a dual prediction consistency loss, and the detection loss includes at least a classification loss and a bounding box regression loss; Based on the joint loss, the parameters of the student model are updated using the backpropagation algorithm to obtain a cross-domain object detection model.

[0011] According to an adaptive object detection method provided by the present invention, the method involves updating the model parameters of the student model based on the joint loss using a backpropagation algorithm to obtain a cross-domain object detection model, comprising: Based on the joint loss, the model parameters of the student model are updated using the backpropagation algorithm to obtain the target student model; Based on the exponential moving average mechanism, the model parameters of the teacher model are updated according to the target student model to obtain the target teacher model; The target teacher model is used as the current teacher model, and the target student model is used as the current student model. The above steps of updating the student model are repeated until the detection performance of the target student model meets the preset convergence condition. The converged target student model is then used as the cross-domain target detection model.

[0012] The present invention also provides an adaptive target detection device, comprising the following modules: The initialization module is used to initialize the student model and the teacher model based on the source domain detection model. The determination module is used to determine a first predicted probability distribution of candidate target boxes based on the original target domain image, using the teacher model, and to determine a second predicted probability distribution of the candidate target boxes using the student model. The input module is used to input the image patch corresponding to the candidate target box into the visual semantic teacher model to obtain the third prediction probability distribution; The determining module is further configured to determine pseudo-labels based on the first predicted probability distribution and the third predicted probability distribution; An update module is used to update the student model based on the pseudo-label and the second predicted probability distribution to obtain a cross-domain target detection model, so as to achieve adaptive target detection based on the cross-domain target detection model.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the adaptive target detection method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the adaptive target detection method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the adaptive target detection method as described above.

[0016] This invention provides an adaptive object detection method, apparatus, electronic device, storage medium, and program product. The method initializes a student model and a teacher model based on a source domain detection model; based on the original target domain image, it determines a first predicted probability distribution of candidate target boxes using the teacher model and a second predicted probability distribution of the candidate target boxes using the student model; it inputs the image blocks corresponding to the candidate target boxes into the visual semantic teacher model to obtain a third predicted probability distribution; it determines pseudo-labels based on the first and third predicted probability distributions; and it updates the student model based on the pseudo-labels and the second predicted probability distribution to obtain a cross-domain object detection model, thereby achieving adaptive object detection based on the cross-domain object detection model. This invention addresses the technical problem of model training instability, which can easily lead to decreased accuracy in the final target detection model when using traditional adaptive techniques. Compared to existing technologies, this invention constructs a dual-teacher supervision system consisting of a source domain teacher model and a visual semantic teacher model. The source domain teacher model ensures the professionalism of pseudo-label detection, while the visual semantic teacher model provides more robust semantic knowledge based on cross-modal generalization capabilities. The pseudo-labels generated by the fusion of these two models are more accurate and reliable, reducing noise accumulation at its source and effectively avoiding the model training instability problem caused by pseudo-label noise accumulation in traditional adaptive techniques. The entire solution requires no additional offline retraining; it can complete real-time adaptive optimization solely based on online data streams from the unlabeled target domain. While ensuring cross-domain adaptation efficiency, it significantly improves the detection accuracy and robustness of cross-domain target detection models under domain shift scenarios such as weather changes, device differences, and sudden scene changes. It can stably adapt to the actual online deployment needs of autonomous driving, security monitoring, and other applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts of the adaptive target detection method provided by the present invention.

[0019] Figure 2 This is a dynamic dual-teaching framework diagram of the adaptive target detection method provided by the present invention.

[0020] Figure 3 This is the second flowchart of the adaptive target detection method provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the adaptive target detection device provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined with Figure 1 and Figure 2 The present invention describes an adaptive target detection method applicable to any adaptive target detection method. The execution subject of this method can be an electronic device or an adaptive target detection device installed in the electronic device. The adaptive target detection device can be implemented by software, hardware, or a combination of both.

[0025] Figure 1 This is one of the flowcharts illustrating the adaptive target detection method provided by this invention, such as... Figure 1 As shown, the method includes the following: Step 101: Based on the source domain detection model, initialize the student model and the teacher model; It should be noted that the source domain detection model refers to an object detection model pre-trained on a labeled source domain dataset; the source domain detection model includes a feature extractor and a detection head. Initializing the student model and teacher model means setting the network structure and initial weight parameters of the student model and teacher model to be exactly the same as those of the source domain detection model.

[0026] It is understandable that the source domain dataset refers to the benchmark dataset used to pre-train object detection models. Such datasets have sufficient annotation information and their data distribution (such as weather conditions, imaging equipment, scene style, etc.) is relatively stable.

[0027] Step 102: Based on the original target domain image, determine the first predicted probability distribution of the candidate target boxes through the teacher model, and determine the second predicted probability distribution of the candidate target boxes through the student model; It should be noted that the original target domain image can be obtained from the target domain test data. This target domain test data is an unlabeled online data stream in the actual deployment scenario of the model (such as autonomous driving, security monitoring, robotics and remote sensing). The distribution of this type of data is often unstable and often faces distribution offset problems such as weather changes, differences in imaging equipment, and sudden changes in scene style, resulting in a significant domain offset from the source domain data.

[0028] In the specific implementation, the original target domain image can be directly input into the teacher model, which performs target detection on the original target domain image and generates candidate target boxes. At the same time, it outputs the corresponding first predicted probability distribution for each candidate target box. The student model uses the same candidate target box detection logic as the teacher model to identify matching candidate target boxes for the same original target domain image, and then outputs the second predicted probability distribution for each candidate target box. This ensures that the teacher and student models predict the probability distribution for the same candidate target box.

[0029] In practical implementation, to construct input differences between teacher and student models to enhance the feature learning generalization of the student model while avoiding the limitations of feature learning caused by a single input scenario, a design of "single-model input augmentation and dual-model input heterogeneity" can be adopted. Specifically, data augmentation processing can be performed only on the original target domain image input to the student model, or only on the original target domain image input to the teacher model, ensuring that there are differences between the images input to the two models. If only the student model input is enhanced, enhancement methods include, but are not limited to, random cropping, horizontal flipping, random brightness / contrast adjustment, and slight color jitter. The teacher model still uses the original target domain image as input. This allows the student model to learn more robust detection capabilities in diverse feature scenarios, while relying on the teacher model's original input benchmark to ensure the scene realism and matching accuracy of candidate target boxes. If only the teacher model input is enhanced, enhancement methods can include random scaling, vertical flipping, Gaussian noise addition, and local blurring. The student model still uses the original target domain image as input. The detection results of the teacher model on the enhanced image provide a more challenging supervisory reference for the student model, prompting the student model to learn adaptability to image variations. Both enhancement strategies ensure the heterogeneity of the input images for the teacher and student models, thus broadening the feature learning boundary of the model through differential inputs and effectively alleviating the adaptation problem caused by the target domain data distribution offset.

[0030] In a specific implementation, taking the target detection scenario of autonomous driving as an example, if the categories to be detected are cars, bicycles and buses, for the candidate target boxes of vehicles on a certain road, the first predicted probability distribution output by the teacher model can be represented as [cars -0.93, bicycles -0.05, buses -0.02]. Subsequent predicted probability distributions can all be represented in the same form.

[0031] Step 103: Input the image patch corresponding to the candidate target box into the visual semantic teacher model to obtain the third prediction probability distribution; It should be noted that the Visual Semantic Teacher Model (VLM) refers to the pre-trained CLIP (Contrastive Language-Image Pre-training) model. It is a visual language model pre-trained on massive image-text pairing data, which has powerful zero-shot recognition capabilities and domain generalization. It can establish the association between visual features and text semantics to perform cross-modal semantic understanding of image patches corresponding to candidate target boxes, thereby generating a third prediction probability distribution.

[0032] Step 104: Determine the pseudo-labels based on the first predicted probability distribution and the third predicted probability distribution; It should be noted that the first predicted probability distribution is generated by the source domain teacher model trained on the source domain dataset, possessing professional object detection capabilities. Its output provides an accurate basis for class judgment of pseudo-labels. However, due to the distribution offset between the source and target domains, relying solely on this distribution is prone to class prediction bias, and the generated labels contain noise interference, making it difficult to directly serve as an effective supervisory signal. The third predicted probability distribution, on the other hand, is generated by a pre-trained visual-semantic teacher model. Leveraging the cross-modal generalization capabilities accumulated from massive image and text data, it effectively mitigates the negative impact of domain offset, accurately corrects the prediction bias of the source domain teacher model, and significantly filters label noise. By adaptively fusing the two probability distributions, the professional detection advantages of the source domain teacher model are retained, while the cross-domain correction capabilities of the visual-semantic teacher model are absorbed. After fusion, the category corresponding to the highest confidence is taken as the pseudo-label, and the bounding box pseudo-label is determined by combining it with the candidate target box parameters output by the source domain teacher model. Finally, a low-noise, high-reliability complete pseudo-label is formed, providing accurate and effective supervisory support for student model training on unlabeled target domain images.

[0033] Step 105: Based on the pseudo-label and the second predicted probability distribution, update the student model to obtain a cross-domain target detection model, so as to achieve adaptive target detection based on the cross-domain target detection model.

[0034] It should be noted that the second prediction probability distribution and the pseudo-label together constitute a closed-loop self-supervised learning framework. The pseudo-label provides high-quality, dual-teacher-verified target supervision signals, while the second prediction probability distribution represents the student model's actual prediction of the original target domain image under the current parameter state. By calculating the differences between the two (such as image-level and target-level classification losses, bounding box regression losses), and supplementing this with a consistency constraint loss to ensure that the student model's predictions remain consistent with the two teachers' predictions, a joint optimization objective driving model updates can be constructed. Based on this objective, by calculating gradients through backpropagation and iteratively updating the student model parameters, its prediction bias on the target domain data can be continuously reduced, gradually improving the model's cross-domain adaptability.

[0035] In specific implementations, such as Figure 2 As shown, the entire cross-domain adaptive object detection process takes unlabeled target domain test data as input and feeds it into the student network and teacher network through two image enhancement paths, strong and weak, respectively. Strong enhancement improves the feature generalization of the student model, while weak enhancement ensures the stability of the detection benchmark of the teacher model. Subsequently, both the student network and the teacher network extract candidate target boxes through ROI pooling, and then output the prediction result P of the student model through RCNN. s Compared with the teacher model prediction result P t Simultaneously, image patches are cropped from the candidate target boxes, and combined with the text-driven input visual semantic teacher model (CLIP), the prediction result P is obtained through cross-modal semantic matching. c ; then P s With P c The system integrates and generates pseudo-labels that are both professional and generalizable, using supervised loss L... st Consistency loss L tc and distillation loss L det The student model is collaboratively optimized, and a closed-loop iteration is formed by updating the EMA of the student model weights. This enables the student model to have a stable and accurate cross-domain adaptive capability for detection on unlabeled target domain data.

[0036] This invention initializes a student model and a teacher model based on a source domain detection model. Based on the original target domain image, the teacher model determines a first predicted probability distribution of candidate target boxes, and the student model determines a second predicted probability distribution of the candidate target boxes. The image blocks corresponding to the candidate target boxes are input into a visual semantic teacher model to obtain a third predicted probability distribution. Based on the first and third predicted probability distributions, pseudo-labels are determined. Based on the pseudo-labels and the second predicted probability distribution, the student model is updated to obtain a cross-domain target detection model, thereby achieving adaptive target detection. This solves the technical problem that traditional adaptive techniques easily lead to model training instability, affecting the accuracy of the final target detection model. Compared to existing technologies, this invention constructs a dual-teacher supervision system of "source domain teacher model + visual semantic teacher model." The source domain teacher model ensures the professionalism of pseudo-label detection, while the visual semantic teacher model provides more robust semantic knowledge based on cross-modal generalization capabilities. The pseudo-labels generated by the fusion of these two models are more accurate and reliable, reducing noise accumulation at the source and effectively avoiding the model training instability problem caused by pseudo-label noise accumulation in traditional adaptive techniques. The entire solution requires no additional offline retraining and can achieve real-time adaptive optimization solely by relying on online data streams from the unlabeled target domain. While ensuring cross-domain adaptation efficiency, it significantly improves the detection accuracy and robustness of the cross-domain target detection model under domain shift scenarios such as weather changes, device differences, and sudden scene changes. It can stably adapt to actual online deployment needs such as autonomous driving and security monitoring.

[0037] Based on any of the above embodiments, the step of determining a first predicted probability distribution of candidate target boxes using the teacher model and a second predicted probability distribution of the candidate target boxes using the student model, based on the original target domain image, includes: Obtain the original target domain image from the target domain test data; The original target domain image is processed to obtain a weakly enhanced image and a strongly enhanced image.

[0038] The weakly enhanced image is input into the teacher model to obtain candidate target boxes and the corresponding first prediction probability distribution; The enhanced image is input into the student model to obtain candidate target boxes and corresponding second prediction probability distributions.

[0039] It should be noted that by performing simple transformation operations on the original target domain image, such as random horizontal flipping, small-angle rotation, or slight brightness adjustment, a weakly enhanced image with less perturbation can be generated. This type of operation preserves the core visual features and semantic information of the original image to the greatest extent. When input into the teacher model, the teacher model can output stable and reliable candidate target boxes and first prediction probability distribution based on low-perturbation, high-fidelity input and output, which can provide a reliable benchmark supervision signal for the learning of the student model.

[0040] It should be noted that complex transformation operations such as color jittering, random grayscale conversion, Gaussian blur, and small-range random cropping can also be performed on the original target domain image to generate a strongly enhanced image with significant perturbation. Inputting this image into the student model can drive the student model to learn robust features under more challenging input conditions, forcing the student model to break away from its dependence on volatile features of the image surface (such as specific lighting, texture, and color) and instead capture the essential semantic and structural features of the target itself.

[0041] The adaptive target detection method provided in this invention establishes a robust supervision-challenge learning mechanism between the teacher and student models by performing differential enhancement processing on the original target domain image. On the one hand, this design ensures that the teacher model can output high-quality and stable supervision signals based on the weakly enhanced image, laying a reliable benchmark for the parameter iteration update of the student model. On the other hand, by introducing diverse perturbations that simulate the domain shift of the real scene through the strongly enhanced image, the student model is driven to extract more essential and generalizable semantic and structural features of the target under more challenging input conditions, thereby effectively alleviating the distribution shift problem in cross-domain detection and significantly improving the robustness and generalization ability of the model in actual deployment scenarios.

[0042] Based on any of the above embodiments, the step of inputting the image patch corresponding to the candidate target box into the visual semantic teacher model to obtain the third prediction probability distribution includes: Obtain integrated text hints for the category to be detected, wherein the integrated text hints include multiple text hints of different styles; The image patch corresponding to the candidate target box is cropped from the original target domain image; The integrated text and the image patch are input into the visual semantic teacher model to obtain the third prediction probability distribution.

[0043] Understandably, the introduction of integrated text prompts aims to enable the visual semantic teacher model to more comprehensively understand the semantic features of target categories in different contexts through multiple text prompts of different styles (such as descriptive, attributed, and contextualized expressions), avoiding semantic bias caused by single text prompts, thereby enhancing the model's generalization ability to target categories. Cropping image patches corresponding to candidate target boxes from the original target domain image focuses on the core visual features of the target region, reduces background noise interference, and allows the visual semantic teacher model to more accurately achieve cross-modal alignment between vision and semantics. Inputting integrated text prompts and image patches into the visual semantic teacher model to generate a third predicted probability distribution essentially involves deep fusion of multimodal information, allowing the model to combine visual features with rich semantic descriptions to correct ambiguities in visual detection. For example, in target categories with similar appearances, the semantic differences in text prompts enable more accurate differentiation, ultimately providing the student model with a supervisory signal that combines visual credibility and semantic consistency, further improving the accuracy and robustness of cross-domain object detection.

[0044] In practical implementation, multi-style integrated text prompts can be generated for each detection category, including descriptive, attribute-based, and contextualized expressions, for different detection scenarios (such as autonomous driving and security monitoring). The text is then standardized and encoded to adapt to the model input. Simultaneously, image patches are cropped based on the pixel coordinates of candidate target boxes, and input into the model after preprocessing such as size normalization and pixel standardization. The visual-semantic teacher model extracts features through both semantic and visual branches, then completes feature interaction through a cross-modal attention fusion layer, finally outputting a third prediction probability distribution based on the fused features. This distribution combines the concreteness of visual features with the abstractness of semantic features, effectively compensating for the judgment bias of single visual detection in domain-offset scenarios, and providing accurate semantic supervision for subsequent multi-supervised signal fusion and student model training.

[0045] The adaptive object detection method provided in this invention introduces a multimodal supervision mechanism that integrates text prompts and a visual semantic teacher model. This mechanism preserves the concreteness of visual detection while achieving accurate semantic supervision. It enriches the semantic dimension of the supervision signal, reduces the risk of misjudgment in domain-biased scenarios of single visual detection, strengthens the alignment capability of cross-modal features to avoid semantic bias interference, and assists the student model in learning more generalized target features. This effectively improves the accuracy and robustness of cross-domain object detection, alleviates the category ambiguity problem of traditional visual detection in distribution-biased scenarios, and enhances the adaptability and detection accuracy of the model in actual deployment scenarios.

[0046] Based on any of the above embodiments, the step of inputting the integrated text and the image patch into the visual semantic teacher model to obtain the third prediction probability distribution includes: The integrated text prompt is input into the text encoder of the visual semantic teacher model to generate the category text features of the category to be detected; The image blocks are input into the image encoder of the visual semantic teacher model to generate image features; Calculate the feature similarity between the image features and the category text features of each category to be detected, and determine the third prediction probability distribution based on the feature similarity.

[0047] In practical implementation, for each category to be detected, multi-style integrated text prompts can be manually designed or automatically generated based on a semantic library, depending on the actual detection scenario (such as autonomous driving, security monitoring, remote sensing recognition, etc.). For example, in the autonomous driving scenario, prompts for the "vehicle" category can be designed as ["a car", "a sedan in a photo", "a vehicle driving on the road"]; in the security monitoring scenario, prompts for the "pedestrian" category can be designed as ["a pedestrian", "a person in the surveillance footage", "a pedestrian walking"]; and in the remote sensing recognition scenario, prompts for the "building" category can be designed as ["a building", "a house in a satellite image", "a structure in an urban area"]. By inputting these text prompts, which are diverse in expression style and contextual emphasis, into the text encoder of the visual semantic teacher model, and aggregating multiple text features of the output (such as average pooling), a more domain-invariant and semantically robust category text feature representation can be generated for each category. This effectively improves the accuracy and stability of the visual language model in zero-shot semantic matching of object categories in open and variable target domains.

[0048] In the specific implementation, for image patches cropped from the original target domain image, they are input into the image encoder of the visual semantic teacher model (such as a Transformer-based visual feature extraction network). Through feature extraction using multi-layer convolution and attention mechanisms, a high-dimensional image feature representation containing key visual details of the target region is generated. This process automatically suppresses background noise interference, ensuring that the image features focus on the visual information of the target itself. Subsequently, the image features are compared with the aggregated text features of each category to be detected using cosine similarity calculation to quantify the degree of matching between visual and semantic features. For example, in autonomous driving scenarios, for image patches within candidate target boxes, the model calculates the cosine similarity between its image features and the aggregated text features of categories such as "vehicle" and "bicycle". The higher the score, the stronger the semantic alignment between the visual features and the corresponding category. Finally, the similarity scores of all categories are input into the Softmax function for normalization, mapping the similarity to a value that conforms to a probability distribution, thus obtaining the third predicted probability distribution.

[0049] Understandably, the third prediction probability distribution represents the category judgment of the target region by the visual semantic teacher model from the perspective of "domain-general semantic knowledge". It complements the judgment of the source domain teacher model based on "domain-specific visual knowledge". It can effectively alleviate the category ambiguity of single visual detection in the domain offset scenario and provide more comprehensive and robust multimodal supervision signals for the student model.

[0050] The adaptive detection method provided in this invention, through the dual-branch feature extraction and cross-modal similarity matching mechanism of the visual semantic teacher model, achieves accurate semantic supervision while preserving the concreteness of visual detection. This not only enriches the semantic dimension of the supervision signal and reduces the risk of misjudgment in domain-biased scenarios of single visual detection, but also strengthens the alignment capability of cross-modal features to avoid semantic bias interference. Furthermore, it can assist the student model in learning more generalized target features, effectively improving the accuracy and robustness of cross-domain target detection, alleviating the category ambiguity problem of traditional visual detection in distribution-biased scenarios, and enhancing the model's adaptability and detection accuracy in practical deployment scenarios such as autonomous driving and security monitoring.

[0051] Based on any of the above embodiments, updating the student model based on the pseudo-label and the second predicted probability distribution to obtain a cross-domain object detection model includes: Based on the pseudo-label and the second predicted probability distribution, a joint loss is determined, wherein the joint loss includes a detection loss and a dual prediction consistency loss, and the detection loss includes at least a classification loss and a bounding box regression loss; Based on the joint loss, the parameters of the student model are updated using the backpropagation algorithm to obtain a cross-domain object detection model.

[0052] It should be noted that the joint loss is the core supervisory signal driving the student model to achieve cross-domain adaptation, and it is formed by fusing the detection loss and the dual prediction consistency loss: where the detection loss L det It includes classification loss and bounding box regression loss. The detection loss constrains the class prediction accuracy and bounding box localization accuracy of the student model; while the dual prediction consistency loss L... con The KL divergence is used to measure the predicted probability distribution P of the student model. s The first predicted probability distribution P of the source domain teacher model t The third prediction probability distribution P of the visual semantic teacher model c The differences are considered, and a dynamic weighting coefficient μ is introduced for weighting (μ=i / N, μ is the fusion coefficient that changes dynamically with the adaptive process, N is the total number of test data in the target domain, and i is the image number of the original target domain image being processed). The specific calculation formula is as follows: Finally, the two types of losses are merged into a joint loss for the student model: It should be noted that the joint loss gradient is propagated back from the model output layer to the feature extraction layer through the backpropagation algorithm, thereby updating the weight parameters of the student model. Under the collaborative supervision of two teachers, the student model gradually acquires cross-domain adaptive detection capabilities, and finally obtains a cross-domain object detection model.

[0053] The adaptive target detection method provided in this invention constructs a joint loss that integrates detection loss and dual prediction consistency loss. This not only ensures the basic detection accuracy of the student model in category prediction and target box localization by relying on the detection loss, but also uses KL divergence to align the prediction distributions of the student model and the dual-teacher model through the dual prediction consistency loss. This enables the student model to learn domain-invariant general features and effectively reduces the sensitivity of detection results to domain shifts. Furthermore, by combining the backpropagation algorithm to update the student model parameters, the model can complete cross-domain adaptive learning without relying on target domain labeled data, significantly improving the model's detection generalization and robustness in unlabeled and variable target domains.

[0054] Based on any of the above embodiments, the step of updating the model parameters of the student model using the backpropagation algorithm based on the joint loss to obtain the cross-domain object detection model includes: Based on the joint loss, the model parameters of the student model are updated using the backpropagation algorithm to obtain the target student model; Based on the exponential moving average mechanism, the model parameters of the teacher model are updated according to the target student model to obtain the target teacher model; The target teacher model is used as the current teacher model, and the target student model is used as the current student model. The above steps of updating the student model are repeated until the detection performance of the target student model meets the preset convergence condition. The converged target student model is then used as the cross-domain target detection model.

[0055] It should be noted that the Exponential Moving Average (EMA) mechanism is the core of ensuring stable updates to the teacher model. Its function is to allow the source domain teacher model to smoothly absorb new knowledge learned by the student model in the target domain, avoiding a sharp drop in teacher model performance due to fluctuations in the student model's training. Specifically, the exponential moving average mechanism is implemented as follows: In the formula, This represents the parameters of the teacher model in the i-th iteration. This represents the parameters of the student model in the (i+1)th iteration. A smoothing coefficient close to 1 (e.g., 0.999) is used to ensure that the updates to the teacher model are gradual and smooth.

[0056] Understandably, by using the updated target teacher model and target student model as the initial models for the next iteration, and cyclically executing the process of "generating predicted probability distribution → determining pseudo-labels → calculating joint loss → updating student model → updating teacher model," the model can continuously and adaptively optimize on unlabeled target domain data streams with varying distributions. When the detection performance of the target student model (such as detection accuracy and loss value on target domain data) meets the preset convergence condition (such as the loss decreasing less than a threshold for multiple consecutive rounds), training terminates. At this point, the converged target student model becomes the final cross-domain object detection model, capable of achieving more stable and accurate detection in the new target domain.

[0057] The adaptive target detection method provided in this invention optimizes the teacher and student models by combining backpropagation of joint loss and exponential moving average (EMA) mechanism. The EMA mechanism smoothly updates the teacher model, avoiding performance fluctuations caused by single-step training fluctuations. Through iterative updates, the model can continuously and adaptively optimize on unlabeled target domain data streams with varying distributions, effectively solving the problem of training instability in traditional adaptive techniques. The resulting cross-domain target detection model has higher detection accuracy and robustness, and can stably cope with various domain offset scenarios.

[0058] Figure 3 This is the second flowchart of the adaptive target detection method provided by the present invention, as shown below. Figure 3 As shown, step 104 also includes steps 1041 to 1042: Step 1041: Determine the fusion probability distribution based on the first predicted probability distribution and the third predicted probability distribution; It should be noted that fusing the first predicted probability distribution output by the teacher model (representing judgments from the perspective of domain-specific visual knowledge) with the third predicted probability distribution output by the visual semantic teacher model (representing judgments from the perspective of domain-general semantic knowledge) can construct a more comprehensive and robust supervision signal.

[0059] In practical implementation, the fusion probability distribution can be directly calculated using an equal-weighted average method, i.e., by using formula P. f =(P t +P c ) / 2 is implemented, where P f For the fusion probability distribution, P t Let P be the first predicted probability distribution. cAs the third prediction probability distribution, this method does not require the design of complex dynamic fusion coefficients and has the advantages of being simple to implement, computationally efficient, and without additional parameter tuning costs.

[0060] Step 1042: When the confidence level corresponding to the fusion probability distribution is greater than a preset threshold, determine the pseudo-label based on the fusion probability distribution.

[0061] It's important to note that after obtaining the fusion probability distribution, the highest probability value in the distribution can be used as the confidence level. Only when this confidence level exceeds a preset threshold (e.g., 0.7) is the category corresponding to the highest probability value determined as a pseudo-label (the pseudo-label includes category information and the coordinate information of the candidate detection box corresponding to that category). This effectively filters out low-confidence predictions that remain questionable after fusion. Specifically, in autonomous driving scenarios, if the highest probability value of the fusion probability distribution corresponding to a candidate target box is 0.6, which does not reach the preset threshold, no pseudo-label will be generated. This avoids noise interference to the training of student models caused by such fuzzy judgments. When the confidence level meets the preset threshold requirement, for example, if the highest probability value of the fusion probability distribution corresponding to a candidate target box is 0.85, then the "vehicle" category corresponding to the highest probability value is determined as a pseudo-label. These pseudo-labels combine the visual detection accuracy of the source domain teacher model with the semantic stability of the visual semantic teacher model, and have undergone rigorous confidence screening. They can provide student models with high-quality supervision signals that are both visually reliable and semantically consistent, thereby effectively reducing the training bias caused by pseudo-label noise in cross-domain scenarios. This allows student models to learn more generalizable target features under reliable supervision, ultimately improving the accuracy and robustness of adaptive object detection.

[0062] The adaptive target detection method provided in this invention employs a pseudo-label generation mechanism of "dual distribution fusion + confidence threshold filtering." This mechanism preserves the richness of the supervision signal while achieving precise control over the quality of pseudo-labels. By fusing the first and third prediction probability distributions, it integrates the accuracy of visual detection with the stability of semantic supervision, improving the credibility and robustness of pseudo-labels. Furthermore, by filtering the fused probability distribution using a preset threshold, it effectively filters out noisy pseudo-labels with low confidence, preventing them from interfering with student model training. It also provides high-quality, high-reliability supervision signals for student models, significantly mitigating training bias caused by pseudo-label noise in cross-domain scenarios. Ultimately, this improves the detection accuracy and generalization ability of the model in real-world scenarios such as autonomous driving and security monitoring.

[0063] Based on any of the above embodiments, determining the fusion probability distribution based on the first predicted probability distribution and the third predicted probability distribution includes: If the highest probability index of the category of the first predicted probability distribution is consistent with that of the third predicted probability distribution, then the first predicted probability distribution is used as the fusion probability distribution. If the highest probability index of the category of the first predicted probability distribution is inconsistent with that of the second predicted probability distribution, then the first predicted probability distribution and the third predicted probability distribution are weighted and fused to obtain a fused probability distribution.

[0064] In the specific implementation, firstly, it is determined whether the prediction results of the first and third predicted probability distributions are consistent: if the highest probability categories predicted by both for the same candidate target box are the same, it indicates that the two teachers' judgments are consistent, and the first predicted probability distribution can be directly used as the fusion probability distribution; if the highest probability categories predicted by both are different, it indicates that there is a cognitive discrepancy, and in this case, a dynamic weighted fusion mechanism needs to be used to calculate the fusion probability distribution. Specifically, the calculation formula of the dynamic weighted fusion mechanism is: P f =(1-μ)P c +μP t , where P t Let P be the first predicted probability distribution. c For the third prediction probability distribution, μ = i / N, where μ is the fusion coefficient that changes dynamically with the adaptive process, N is the total number of test data in the target domain, and i is the image number of the original target domain image being processed.

[0065] Understandably, in the initial stage of adaptation, the source domain teacher is severely affected by domain drift, and its predictions are unreliable. At this time, more trust should be placed on the language-driven teacher with strong generalization ability. As adaptation progresses, the source domain teacher updates its knowledge of the target domain, and its predictive ability gradually improves. At this point, trust in it should be gradually increased. Specifically, as the adaptation process advances, i.e., the image index i being processed continuously increases from 1, the value of the fusion coefficient μ will also gradually increase linearly from close to 0 to close to 1. This change directly determines the weight allocation of the above-mentioned dynamic weighted fusion mechanism: in the initial stage of adaptation (μ is very small), the fusion result P... f The prediction P is highly dependent on language-driven teachers with greater domain generalization (i.e., visual semantic teacher models). c As the process progresses (μ increases), the predicted P of the source domain teachers (i.e., the teacher model) that have been gradually adapted to the target domain through updates... t Its weight in integration is constantly increasing.

[0066] The adaptive target detection method provided in this invention adopts a probability distribution fusion strategy of "consistency priority and divergence complementarity" to achieve dynamic enhancement of the supervision signal while preserving the visual detection accuracy of the source domain teacher model. This avoids redundant calculations and improves training efficiency. Furthermore, the weighted fusion in divergence scenarios strengthens the robustness of supervision to avoid single judgment bias. It can also rely on the complementary advantages of dual teacher models to provide a supervision benchmark for student models that has both visual accuracy and semantic stability. This effectively alleviates the category ambiguity problem of traditional cross-domain detection in distribution offset scenarios and improves the detection accuracy and generalization ability of the model in actual deployment scenarios such as autonomous driving and security monitoring.

[0067] The adaptive target detection device provided by the present invention is described below. The adaptive target detection device described below can be referred to in correspondence with the adaptive target detection method described above. Figure 4 As shown, the adaptive target detection device includes: Initialization module 10 is used to initialize the student model and the teacher model based on the source domain detection model; The determination module 20 is used to determine a first predicted probability distribution of candidate target boxes based on the original target domain image, through the teacher model, and to determine a second predicted probability distribution of the candidate target boxes through the student model. Input module 30 is used to input the image patch corresponding to the candidate target box into the visual semantic teacher model to obtain the third prediction probability distribution; The determining module 20 is further configured to determine pseudo-labels based on the first predicted probability distribution and the third predicted probability distribution; The update module 40 is used to update the student model based on the pseudo-label and the second predicted probability distribution to obtain a cross-domain target detection model, so as to achieve adaptive target detection based on the cross-domain target detection model.

[0068] Optionally, the determining module 20 is further configured to: Obtain the original target domain image from the target domain test data; The original target domain image is processed to obtain a weakly enhanced image and a strongly enhanced image; The weakly enhanced image is input into the teacher model to obtain candidate target boxes and the corresponding first prediction probability distribution; The enhanced image is input into the student model to obtain candidate target boxes and corresponding second prediction probability distributions.

[0069] Optionally, the input module 30 is further configured to: Obtain integrated text hints for the category to be detected, wherein the integrated text hints include multiple text hints of different styles; The image patch corresponding to the candidate target box is cropped from the original target domain image; The integrated text and the image patch are input into the visual semantic teacher model to obtain the third prediction probability distribution.

[0070] Optionally, the input module 30 is further configured to: The integrated text prompt is input into the text encoder of the visual semantic teacher model to generate the category text features of the category to be detected; The image blocks are input into the image encoder of the visual semantic teacher model to generate image features; Calculate the feature similarity between the image features and the category text features of each category to be detected, and determine the third prediction probability distribution based on the feature similarity.

[0071] Optionally, the determining module 20 is further configured to: Based on the first predicted probability distribution and the third predicted probability distribution, a fusion probability distribution is determined; When the confidence level corresponding to the fusion probability distribution is determined to be greater than a preset threshold, a pseudo-label is determined based on the fusion probability distribution.

[0072] Optionally, the determining module 20 is further configured to: If the highest probability index of the category of the first predicted probability distribution is consistent with that of the third predicted probability distribution, then the first predicted probability distribution is used as the fusion probability distribution. If the highest probability index of the category of the first predicted probability distribution is inconsistent with that of the second predicted probability distribution, then the first predicted probability distribution and the third predicted probability distribution are weighted and fused to obtain a fused probability distribution.

[0073] Optionally, the update module 40 is further configured to: Based on the pseudo-label and the second predicted probability distribution, a joint loss is determined, wherein the joint loss includes a detection loss and a dual prediction consistency loss, and the detection loss includes at least a classification loss and a bounding box regression loss; Based on the joint loss, the model parameters of the student model are updated using the backpropagation algorithm to obtain a cross-domain object detection model.

[0074] Optionally, the update module 40 is further configured to: Based on the joint loss, the model parameters of the student model are updated using the backpropagation algorithm to obtain the target student model; Based on the exponential moving average mechanism, the model parameters of the teacher model are updated according to the target student model to obtain the target teacher model; The target teacher model is used as the current teacher model, and the target student model is used as the current student model. The above steps of updating the student model are repeated until the detection performance of the target student model meets the preset convergence condition. The converged target student model is then used as the cross-domain target detection model.

[0075] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute an adaptive object detection method. This method includes: initializing a student model and a teacher model based on a source domain detection model; determining a first predicted probability distribution of candidate target boxes based on the original target domain image using the teacher model, and determining a second predicted probability distribution of the candidate target boxes using the student model; inputting the image blocks corresponding to the candidate target boxes into the visual semantic teacher model to obtain a third predicted probability distribution; determining pseudo-labels based on the first and third predicted probability distributions; and updating the student model based on the pseudo-labels and the second predicted probability distribution to obtain a cross-domain object detection model, thereby achieving adaptive object detection based on the cross-domain object detection model.

[0076] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the adaptive object detection method provided by the above methods. The method includes: initializing a student model and a teacher model based on a source domain detection model; determining a first predicted probability distribution of candidate target boxes based on the teacher model and a second predicted probability distribution of the candidate target boxes based on the student model based on the original target domain image; inputting the image blocks corresponding to the candidate target boxes into a visual semantic teacher model to obtain a third predicted probability distribution; determining pseudo-labels based on the first and third predicted probability distributions; and updating the student model based on the pseudo-labels and the second predicted probability distribution to obtain a cross-domain object detection model, thereby achieving adaptive object detection based on the cross-domain object detection model.

[0078] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the adaptive object detection method provided by the methods described above. This method includes: initializing a student model and a teacher model based on a source domain detection model; determining a first predicted probability distribution of candidate target boxes using the teacher model based on an original target domain image, and determining a second predicted probability distribution of the candidate target boxes using the student model; inputting image blocks corresponding to the candidate target boxes into a visual semantic teacher model to obtain a third predicted probability distribution; determining pseudo-labels based on the first and third predicted probability distributions; and updating the student model based on the pseudo-labels and the second predicted probability distribution to obtain a cross-domain object detection model, thereby achieving adaptive object detection based on the cross-domain object detection model.

[0079] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0080] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive target detection method, characterized in that, include: Based on the source domain detection model, initialize the student model and the teacher model; Based on the original target domain image, the teacher model determines the first predicted probability distribution of the candidate target boxes, and the student model determines the second predicted probability distribution of the candidate target boxes. The image patch corresponding to the candidate target box is input into the visual semantic teacher model to obtain the third prediction probability distribution; Based on the first predicted probability distribution and the third predicted probability distribution, pseudo-labels are determined; Based on the pseudo-labels and the second predicted probability distribution, the student model is updated to obtain a cross-domain target detection model, so as to achieve adaptive target detection based on the cross-domain target detection model.

2. The adaptive target detection method according to claim 1, characterized in that, The process of determining a first predicted probability distribution of candidate target boxes based on the original target domain image using the teacher model and a second predicted probability distribution of the candidate target boxes using the student model includes: Obtain the original target domain image from the target domain test data; The original target domain image is processed to obtain a weakly enhanced image and a strongly enhanced image; The weakly enhanced image is input into the teacher model to obtain candidate target boxes and the corresponding first prediction probability distribution; The enhanced image is input into the student model to obtain candidate target boxes and corresponding second prediction probability distributions.

3. The adaptive target detection method according to claim 1, characterized in that, The step of inputting the image patch corresponding to the candidate target box into the visual semantic teacher model to obtain the third prediction probability distribution includes: Obtain integrated text hints for the category to be detected, wherein the integrated text hints include multiple text hints of different styles; The image patch corresponding to the candidate target box is cropped from the original target domain image; The integrated text and the image patch are input into the visual semantic teacher model to obtain the third prediction probability distribution.

4. The adaptive target detection method according to claim 3, characterized in that, The step of inputting the integrated text and the image patch into the visual semantic teacher model to obtain the third predicted probability distribution includes: The integrated text prompt is input into the text encoder of the visual semantic teacher model to generate the category text features of the category to be detected; The image blocks are input into the image encoder of the visual semantic teacher model to generate image features; Calculate the feature similarity between the image features and the category text features of each category to be detected, and determine the third prediction probability distribution based on the feature similarity.

5. The adaptive target detection method according to claim 1, characterized in that, The step of determining pseudo-labels based on the first predicted probability distribution and the third predicted probability distribution includes: Based on the first predicted probability distribution and the third predicted probability distribution, a fusion probability distribution is determined; When the confidence level corresponding to the fusion probability distribution is determined to be greater than a preset threshold, a pseudo-label is determined based on the fusion probability distribution.

6. The adaptive target detection method according to claim 5, characterized in that, Determining the fusion probability distribution based on the first predicted probability distribution and the third predicted probability distribution includes: If the highest probability index of the category of the first predicted probability distribution is consistent with that of the third predicted probability distribution, then the first predicted probability distribution is used as the fusion probability distribution. If the highest probability index of the category of the first predicted probability distribution is inconsistent with that of the second predicted probability distribution, then the first predicted probability distribution and the third predicted probability distribution are weighted and fused to obtain a fused probability distribution.

7. The adaptive target detection method according to claim 1, characterized in that, The step of updating the student model based on the pseudo-labels and the second predicted probability distribution to obtain a cross-domain object detection model includes: Based on the pseudo-label and the second predicted probability distribution, a joint loss is determined, wherein the joint loss includes a detection loss and a dual prediction consistency loss, and the detection loss includes at least a classification loss and a bounding box regression loss; Based on the joint loss, the model parameters of the student model are updated using the backpropagation algorithm to obtain a cross-domain object detection model.

8. The adaptive target detection method according to claim 7, characterized in that, The process of updating the model parameters of the student model based on the joint loss using the backpropagation algorithm to obtain a cross-domain object detection model includes: Based on the joint loss, the model parameters of the student model are updated using the backpropagation algorithm to obtain the target student model; Based on the exponential moving average mechanism, the model parameters of the teacher model are updated according to the target student model to obtain the target teacher model; The target teacher model is used as the current teacher model, and the target student model is used as the current student model. The above steps of updating the student model are repeated until the detection performance of the target student model meets the preset convergence condition. The converged target student model is then used as the cross-domain target detection model.

9. An adaptive target detection device, characterized in that, include: The initialization module is used to initialize the student model and the teacher model based on the source domain detection model. The determination module is used to determine a first predicted probability distribution of candidate target boxes based on the original target domain image, using the teacher model, and to determine a second predicted probability distribution of the candidate target boxes using the student model. The input module is used to input the image patch corresponding to the candidate target box into the visual semantic teacher model to obtain the third prediction probability distribution; The determining module is further configured to determine pseudo-labels based on the first predicted probability distribution and the third predicted probability distribution; An update module is used to update the student model based on the pseudo-label and the second predicted probability distribution to obtain a cross-domain target detection model, so as to achieve adaptive target detection based on the cross-domain target detection model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the adaptive target detection method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the adaptive target detection method as described in any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the adaptive target detection method as described in any one of claims 1 to 8.

Images