Semi-supervised SAR ship target detection method based on pseudo labeling and active learning
By combining saliency pseudo-labeling and teacher-student consistency semi-supervised training with dual uncertainty active learning, the problems of high-cost labeling and unstable pseudo-labeling in SAR ship detection are solved, achieving high-precision target detection, which is suitable for maritime surveillance and marine situational awareness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-07
- Publication Date
- 2026-04-03
AI Technical Summary
Existing deep learning object detection methods rely on a large amount of accurately labeled data in SAR images. Manual labeling is costly and unlabeled data is difficult to utilize. The quality of pseudo-labels is unstable, which affects the stability of the model. Active learning is easily affected by noise and background interference in SAR ship detection, resulting in a waste of labeling resources.
A collaborative approach combining saliency pseudo-labeling, teacher-student consistency semi-supervised training, and active sampling with dual uncertainties is adopted. Initial supervision signals are provided through saliency pseudo-labeling, teacher pseudo-labels are generated by the teacher network to constrain the training of the student network, and high-value unlabeled samples are screened for manual labeling through dual uncertainties, forming a closed-loop iterative process.
High-precision SAR ship target detection is achieved with low annotation cost, suppressing the accumulation of false annotation noise, improving model stability and robustness, effectively utilizing unlabeled data, and reducing the waste of annotation resources.
Smart Images
Figure CN121784738A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target detection technology, specifically relating to a semi-supervised SAR ship target detection method based on pseudo-labeling and active learning. Background Technology
[0002] Synthetic Aperture Radar (SAR) has all-weather, all-day imaging capabilities, making it suitable for scenarios such as maritime traffic monitoring, maritime surveillance, and marine situational awareness. However, due to the characteristics of SAR images, such as strong speckle noise, complex background scattering, and significant interference from shoreline and port structures, the distinguishability between ship targets and the background is low, resulting in greater detection difficulty.
[0003] Existing deep learning object detection methods typically rely on a large amount of accurately labeled data. In SAR object detection scenarios, manual labeling not only requires providing the target location and bounding box, but also requires combining scattering characteristics to determine the authenticity of the target. Labeling costs are high and the cycle is long. In practical applications, only a small proportion of labeled data can be obtained, while a large number of images remain unlabeled. Unlabeled data is difficult to fully utilize, thus limiting detection performance.
[0004] Semi-supervised object detection methods typically utilize unlabeled data through "pseudo-labels + consistency constraints," but the quality of pseudo-labels determines the upper limit of training. In SAR object detection scenarios, noise and complex backgrounds can easily cause false detections and false negatives. Incorrect pseudo-labels may accumulate during iterative training, thus affecting model stability.
[0005] Active learning, by manually labeling the most valuable samples from unlabeled samples, can achieve greater performance improvements with the same labeling budget. However, existing uncertainty sampling or diversity sampling is mostly designed for natural image tasks. In SAR ship detection, it is easily affected by strong scattering noise and background structure, causing sampling to favor unlabeled samples with "complex backgrounds" rather than "rich target information", resulting in a waste of labeling resources. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention aims to solve the technical problem of proposing a semi-supervised SAR ship target detection method based on pseudo-labeling and active learning. Through the synergistic effect of saliency pseudo-labeling, teacher-student consistency semi-supervised training, and active sampling with dual uncertainties, high-precision ship target detection is achieved with low labeling costs.
[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A semi-supervised SAR ship target detection method based on pseudo-labeling and active learning is characterized by the following steps: Step 1: Obtain a small number of manually labeled SAR images and a large number of unlabeled SAR images to construct labeled SAR image datasets and unlabeled SAR image datasets; Step 2: Randomly sample some unlabeled SAR images and perform saliency pseudo-labeling to obtain saliency pseudo-labeled SAR images, and add the saliency pseudo-labeled SAR images to the labeled SAR image dataset; For pixels in unlabeled SAR images, a multi-scale sliding window is used to obtain the neighborhood, and the pixel amplitude within the sliding window is calculated. The pixel amplitudes within the multi-scale sliding window are then fused to obtain the brightness saliency feature of the pixel. All pixels are traversed to obtain the brightness saliency feature of the unlabeled SAR image. The gradient operator is used to calculate the mean gradient response of the unlabeled SAR image under the multi-scale sliding window, and the mean gradient response under the multi-scale sliding window is fused to obtain the edge saliency feature. The local texture statistics of the unlabeled SAR image under the multi-scale sliding window are calculated, and the local texture statistics under the multi-scale sliding window are fused to obtain the texture saliency feature. After normalizing the brightness saliency feature, edge saliency feature, and texture saliency feature, a weighted fusion is performed to obtain the saliency map. The saliency map is binarized, and then connected component analysis is performed to obtain multiple connected components. The area and aspect ratio of each connected component are calculated. Connected components whose area or aspect ratio is outside the preset range are classified as sparse connected components and removed, while those that are within the preset range are classified as dense connected components and retained. The DBSCAN clustering algorithm is used to cluster the pixels in the dense connected components to generate multiple pixel clusters. The minimum bounding rectangle of the pixel cluster is used as the bounding box, and the saliency response intensity is used as the confidence level to obtain the saliency pseudo-labeled bounding box. All dense connected components are traversed to obtain the saliency pseudo-labeled SAR image. Step 3: Construct a target detection model, perform consistent semi-supervised training on the target detection model based on a teacher-student network architecture, and use the trained target detection model for SAR ship target detection. The teacher network is pre-trained using an annotated SAR image dataset, the student network is initialized based on the parameters of the pre-trained teacher network, and the student network is then trained under supervision using the annotated SAR image dataset. Weak and strong enhancements are performed on the same unlabeled SAR image to obtain weakly enhanced and strongly enhanced unlabeled SAR images. The weakly enhanced unlabeled SAR image is passed to a pre-trained teacher network to obtain the predicted category, predicted bounding box, and predicted confidence. Predicted bounding boxes with a prediction confidence higher than a threshold are used as teacher pseudo-labels to constrain the unsupervised training of the student network. The strongly enhanced unlabeled SAR image is input into the student network for unsupervised training. During the semi-supervised training process, unlabeled SAR images are sampled from the unlabeled SAR image dataset through active learning of double uncertainty and manually labeled. The manually labeled SAR images are then added to the labeled SAR image dataset. The unlabeled SAR image dataset and the labeled SAR image dataset are updated, and the updated datasets participate in the subsequent training process. Dual uncertainty active learning sampling includes the following: The unlabeled SAR image is processed by the currently trained teacher network to generate candidate boxes, category prediction probabilities, and prediction confidence scores; the detection difficulty uncertainty score of the unlabeled SAR image is calculated according to equation (3): (3) In the formula, Indicates the first Uncertainty score for the detection difficulty of unlabeled SAR images. Indicates the first In the unlabeled SAR image, the first The predicted class probability of each candidate box; The information uncertainty score of the unlabeled SAR image is calculated according to equation (4): (4) In the formula, Indicates the first Uncertainty score of information content in unlabeled SAR images. Indicates the first In the unlabeled SAR image, the first The maximum prediction confidence of each candidate box; The uncertainty scores for detection difficulty and information content are normalized and then summed to obtain the comprehensive uncertainty score for the unlabeled SAR image. The unlabeled SAR image dataset is traversed to obtain the comprehensive uncertainty score of each unlabeled SAR image; several unlabeled SAR images with high comprehensive uncertainty scores are selected for manual annotation.
[0008] Compared with the prior art, the present invention has the following beneficial effects: 1. By performing saliency pseudo-annotations on unlabeled SAR images and using them as initial data for supervised training, a cold start for network training with little or no manually labeled images is achieved. During the semi-supervised training phase, a pre-trained teacher network generates teacher pseudo-annotations on the unlabeled SAR images to constrain the prediction consistency of the student network during unsupervised training, and these are continuously updated with iterations. The saliency pseudo-annotations can be considered the initial source of pseudo-annotations for the teacher network's pre-training, while the teacher pseudo-annotations gradually replace and clean up the saliency pseudo-annotations during training, thereby suppressing the accumulation of pseudo-annotation noise.
[0009] 2. In the semi-supervised training phase, different augmentation strategies are applied to the same unlabeled SAR image. The weakly augmented and strongly augmented unlabeled SAR images are input into the teacher network and the student network, respectively. The predicted bounding boxes output by the teacher network are selected as teacher pseudo-labels based on the confidence level. This is used to constrain the prediction consistency of the student network on the strongly augmented unlabeled SAR image, thus balancing the stability of the teacher pseudo-label predictions with the robustness of the student network training and reducing prediction perturbations.
[0010] 3. In active learning sampling, unlabeled SAR images that are both difficult and informative are selected for manual annotation based on a dual uncertainty metric. This further improves model performance while reducing annotation costs. Both pseudo-annotation and active learning sampling target unlabeled SAR images. The goal of active learning sampling is to "increase high-value manual annotations," while the goal of pseudo-annotation is to "provide training supervision signals based on unlabeled SAR images at low cost." The two complement each other; pseudo-annotation expands the available supervision scale, while active learning sampling improves the quality and efficiency of limited manual annotations. Attached Figure Description
[0011] Figure 1 This is an overall framework diagram of the present invention; Figure 2 This is a flowchart of the uncertainty score calculation process of the present invention; Figure 3 This is a visualization of the detection results of the present invention in strong clutter in the open sea and densely populated near-shore scenes. Detailed Implementation
[0012] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail and are not intended to limit the scope of protection of this application.
[0013] like Figure 1-3 As shown, this invention provides a semi-supervised SAR ship target detection method based on pseudo-labeling and active learning, comprising the following steps: Step 1: Obtain a small number of manually labeled SAR images and a large number of unlabeled SAR images to construct labeled SAR image datasets and unlabeled SAR image datasets; Step 2: Randomly sample some unlabeled SAR images and perform saliency pseudo-labeling to provide spatial prior constraints on potential target areas without relying on the detection model, resulting in saliency pseudo-labeled SAR images; add the saliency pseudo-labeled SAR images to the labeled SAR image dataset to expand the labeled SAR image dataset for cold start of semi-supervised training. (1) Constructing a saliency map To address the significant differences between ship targets in SAR images and the sea surface background in terms of reflection intensity, structural outline, and scattering texture, brightness saliency, edge saliency, and texture saliency features were constructed for unlabeled SAR images.
[0014] For any pixel in an unlabeled SAR image, a multi-scale sliding window is used to obtain its neighborhood, and the pixel amplitude within the sliding window (i.e., the amplitude quantization result of the synthetic aperture radar echo signal) is calculated to highlight the high backscattering bright spot area corresponding to the ship. The pixel amplitudes within the multi-scale sliding window are fused to obtain the brightness saliency feature of the pixel. All pixels are traversed to obtain the brightness saliency feature of the unlabeled SAR image. The average gradient response of unlabeled SAR images under a multi-scale sliding window is calculated using gradient operators such as Sobel. The average gradient response under the multi-scale sliding window is then fused to obtain edge saliency features to characterize the ship structure outline. The LBP algorithm is used to calculate the local texture statistics of unlabeled SAR images under a multi-scale sliding window. The local texture statistics under the multi-scale sliding window are fused to obtain texture saliency features, which characterize the difference in scattering texture between the ship target and the sea surface background. The saliency features, edge saliency features, and texture saliency features are normalized. The normalized saliency features, edge saliency features, and texture saliency features are then weighted and fused to obtain a saliency map.
[0015] In this embodiment, the scales of the multi-scale sliding window are 3×3, 7×7 and 11×11 pixels, respectively.
[0016] (2) Connected component filtering An adaptive threshold (0.42 in this embodiment) is used to perform binarization segmentation on the saliency map to obtain a binarized saliency map. Connectivity analysis is performed on the binarized saliency map to obtain multiple connected components. The area and aspect ratio of each connected component are calculated. Connective components whose area or aspect ratio is not within the preset range are regarded as sparse connected components and are removed. Conversely, they are regarded as dense connected components and are retained to reduce the influence of fragmented noise and obvious non-ship areas.
[0017] (3) Dense Connectivity Clustering and Pseudo-labeling The DBSCAN clustering algorithm is used to cluster pixels within densely connected regions, generating multiple pixel clusters. Each pixel cluster corresponds to a ship target or an interference-free local target region, eliminating redundant regions caused by strong scattering from the coastline and continuous clutter, ensuring that each pixel cluster contains only a single ship target. The minimum bounding rectangle of the pixel cluster is used as the bounding box, and the saliency response intensity (calculated from the saliency features of the corresponding region in the saliency map) is used as the confidence level to obtain saliency pseudo-labeled bounding boxes. All densely connected regions are traversed to obtain saliency pseudo-labeled SAR images. For densely connected regions with large areas (e.g., areas greater than 500 pixels), to avoid excessively large bounding boxes caused by strong scattering from the coastline or clutter, the neighborhood radius is set to 30 and the minimum number of points is 5 during clustering; for densely connected regions with small areas (e.g., areas less than or equal to 500 pixels), the neighborhood radius is set to 5 and the minimum number of points is 2 during clustering.
[0018] Step 3: Construct a target detection model, perform consistent semi-supervised training on the target detection model based on a teacher-student network architecture, and use the trained target detection model for SAR ship target detection. (1) Constructing an object detection model A two-stage object detection model, such as the Faster R-CNN model, is selected, which includes a backbone feature extraction network, a feature pyramid network, a region candidate network, and a detection head. The backbone feature extraction network is used to extract multi-scale features from the input image, the feature pyramid network is used to fuse multi-scale features to enhance the detection capability of small objects, the region candidate network is used to generate candidate regions, and the detection head is used to classify objects and regress bounding boxes in the candidate regions.
[0019] (2) The teacher network is pre-trained and the pre-trained teacher network is used to generate teacher pseudo-labels for unlabeled SAR images to constrain the unsupervised training of the student network. The target detection model is used as the teacher network and the student network respectively. The teacher network is pre-trained using an annotated SAR image dataset to obtain the pre-trained teacher network. The student network is then initialized based on the parameters of the pre-trained teacher network.
[0020] To balance the stability of teacher-labeled predictions with the robustness of student network training, the same unlabeled SAR image is subjected to both weak and strong enhancement to reduce prediction perturbations, resulting in weakly enhanced and strongly enhanced unlabeled SAR images. The weakly enhanced unlabeled SAR image serves as the input image for the pre-trained teacher network, while the strongly enhanced unlabeled SAR image serves as the input image for the student network. Weak enhancement includes one of random flipping or rotation, while strong enhancement includes one or more of random horizontal flipping, color perturbation, Gaussian blurring, and random occlusion.
[0021] During semi-supervised training, weakly enhanced unlabeled SAR images are input into a pre-trained teacher network to obtain predicted categories, predicted bounding boxes, and predicted confidence scores. Non-maximum suppression is applied to the predicted bounding boxes output by the pre-trained teacher network, and the predicted bounding boxes are filtered based on their predicted confidence scores. Predicted bounding boxes with predicted confidence scores higher than a threshold are used as teacher pseudo-labels. In this embodiment, the threshold for predicted confidence scores is 0.7.
[0022] (3) Semi-supervised training of students' networks Supervised training of the student network is performed using an annotated SAR image dataset, and supervised loss is calculated. Unsupervised training of the student network on strongly enhanced unannotated SAR images is then constrained by teacher pseudo-annotations. Specifically, the student network outputs predicted bounding boxes from strongly enhanced unannotated SAR images, and these predicted boxes are compared with the teacher pseudo-annotations to calculate the unsupervised consistency loss, thereby guiding the student network to learn stable and reliable target representations. The total training loss of the student network includes both supervised and unsupervised consistency losses; therefore, the loss function is: (1) In the formula, This represents the total training loss of the student network; This indicates a loss of oversight. This represents the loss of unsupervised consistency. This represents the weighting coefficient, which is set to 6 in this embodiment; To ensure the consistency and reliability of teacher pseudo-labels during training, an exponential moving average method is used to update the teacher network parameters based on the student network parameters. The update formula is as follows: (2) In the formula, , This indicates the teacher network parameters after the update compared to the previous version. Indicates student network parameters; This represents the smoothing coefficient, which takes a value close to 1; in this embodiment, it is set to 0.9996.
[0023] Through the above update method, the teacher network can smoothly track changes in student network parameters, effectively suppress the interference of noise gradients on teacher pseudo-labels, and thus improve the quality of teacher pseudo-labels.
[0024] (4) Active learning sampling with double uncertainty As training progresses, while the teacher network can generate relatively stable pseudo-labels based on unlabeled SAR images, it may still exhibit high prediction uncertainty in some complex scenes or areas with dense targets. To further improve the utilization efficiency of limited manual labeling resources, manual labels are prioritized for SAR images that are most critical to model improvement. An active learning sampling strategy is introduced to evaluate and filter the value of unlabeled SAR images and update the labeled SAR image dataset.
[0025] The unlabeled SAR image is processed by the currently trained teacher network to generate candidate boxes, category prediction probabilities, and prediction confidence. Based on these candidate boxes, image-level uncertainty is modeled from two complementary perspectives: detection difficulty and information content. Detection difficulty uncertainty score and information content uncertainty score are calculated respectively.
[0026] The entropy value is calculated based on the category prediction probability of each candidate box, resulting in a detection difficulty uncertainty score for unlabeled SAR images. A higher score indicates greater detection difficulty uncertainty, meaning the model is less certain about classifying targets in unlabeled SAR images, and unlabeled SAR images present a greater learning challenge for the currently trained model. The formula for calculating the detection difficulty uncertainty score is: (3) In the formula, Indicates the first Uncertainty score for the detection difficulty of unlabeled SAR images. Indicates the first In the unlabeled SAR image, the first The predicted class probability of each candidate box; To further measure the richness of potential targets in an image, information uncertainty is introduced as a supplementary metric. The maximum prediction confidence of each candidate box is accumulated to obtain an information uncertainty score. A higher information uncertainty score indicates that the image may contain more potential targets or have a more complex target distribution, thus possessing higher annotation value. The information uncertainty score is calculated as follows: (4) In the formula, Indicates the first Uncertainty score of information content in unlabeled SAR images. Indicates the first The maximum prediction confidence of each candidate box; The uncertainty scores for detection difficulty and information content are normalized and then summed to obtain the comprehensive uncertainty score for the unlabeled SAR image. (5) In the formula, Indicates the first The overall uncertainty score of Zhang's unlabeled SAR image. , Indicates the first Uncertainty scores for normalized detection difficulty and information content in unlabeled SAR images; The higher the overall uncertainty score, the more difficult it is for the currently trained model to discriminate unlabeled SAR images, while they also contain richer target information, making them more suitable as priority objects for precise manual annotation.
[0027] The unlabeled SAR image dataset is traversed to obtain the comprehensive uncertainty score of each unlabeled SAR image. Based on the comprehensive uncertainty score, all unlabeled SAR images are sorted in descending order. Several unlabeled SAR images with high comprehensive uncertainty scores (e.g., accounting for 2% to 5% of the unlabeled SAR image dataset) are manually labeled. The selected unlabeled SAR images are removed from the unlabeled SAR image dataset, and the manually labeled SAR images are added to the labeled SAR image dataset. The unlabeled SAR image dataset and the labeled SAR image dataset are updated. The updated two datasets re-participate in the subsequent semi-supervised training process. At the same time, the teacher pseudo-labels are updated based on the updated unlabeled SAR image dataset, thus forming a closed-loop iterative process of "semi-supervised training - active sampling - label set expansion - pseudo-label update".
[0028] In this embodiment, active sampling can be performed every 5 training rounds, for a total of 2 times. It is important to emphasize that the core purpose of this iterative process is to simultaneously optimize the target detection model parameters and improve the quality of the supervision signal: on the one hand, the model's detection performance is continuously improved by utilizing unlabeled SAR images through a consistent semi-supervised learning mechanism; on the other hand, the limited manual annotation resources are concentrated on the SAR images that contribute the most to the improvement of model performance through an active learning strategy, thereby achieving higher detection accuracy with lower manual annotation costs.
[0029] When the preset maximum number of iterations is reached, or when the improvement in detection performance is less than a preset threshold in several consecutive evaluation rounds, the iterative training process is terminated; the trained teacher network is then used as the trained target detection model to obtain more stable and reliable detection results.
[0030] Example To verify the effectiveness of the method of this invention, this embodiment conducts an experimental evaluation on the SAR ship detection dataset LS-SSDD-v1.0. This dataset was acquired by the Sentinel-1 satellite and contains 15 large-scene SAR images with a resolution of 24000×16000 pixels, labeled with 6015 ship targets.
[0031] During the saliency pseudo-labeling process, the preset range for the area of connected components is 20-1000, and the preset range for the aspect ratio of connected components is 0.3-3.2. When the area of connected components is greater than 500 pixels, the neighborhood radius of the DBSCAN clustering algorithm is 30 and the minimum number of points is 5. When the area of connected components is less than or equal to 500 pixels, the neighborhood radius of the DBSCAN clustering algorithm is 5 and the minimum number of points is 2.
[0032] In the semi-supervised training phase, there were 2000 pre-training epochs and a total of 180,000 training epochs. For the first 1000 epochs, the learning rate was linearly increased from 0.001 to 0.01 and then kept constant. The pseudo-label confidence threshold was set to 0.7, the sliding coefficient to 0.9996, the weight of the unsupervised consistency loss to 6, and the batch size to 8, with labeled and unlabeled samples each comprising half. In the active learning phase, the number of active sampling iterations was set to 2. Half of the labeled SAR image dataset was generated by salient pseudo-labels, and the other half was selected by uncertain active sampling.
[0033] To comprehensively evaluate the performance of the method of this invention, comparative experiments were conducted. The baseline method only uses saliency pseudo-labels for supervised training. U-ShipDet-Ⅰ refers to a method that uses its own pseudo-labels for semi-supervised consistency training. U-shipdet-Ⅱ refers to a method that randomly selects 5% of unlabeled data for manual labeling based on U-ShipDet-Ⅰ. This application uses uncertainty active learning to sample 5% of unlabeled data for manual labeling for training. The experimental results of each method are shown in Table 1.
[0034] Table 1. Experimental results of different methods
[0035] As shown in Table 1, the method of the present invention significantly improves all indicators under the same annotation budget. Therefore, under limited annotation conditions, the performance of the method of the present invention is significantly better than that of the comparative method.
[0036] Figure 3It is evident that in the scenario of strong clutter in the open sea, the method of the present invention can effectively suppress false detections caused by high backscattering waves and reefs; in the dense ship area near the shore, it can distinguish adjacent targets and reduce target merging; in complex scenarios such as narrow inland rivers, it can still stably detect small-scale ship targets, demonstrating good generalization ability and scene adaptability.
[0037] Figure 3 (a) and Figure 3 (b) shows the detection results of traditional target detection methods (referring to target detection methods that do not introduce saliency annotation, teacher-student consistency semi-supervised training, and dual uncertainty active sampling) and the method of the present invention in a strong cluttered ocean scene. In this type of scene, there is strong backscattering noise and local strong bright spot interference on the sea surface. Traditional target detection methods are prone to misidentifying non-ship targets such as waves or reefs as ship targets. However, the method of the present invention can effectively suppress the above interference and generate detection boxes only at the actual ship target locations, thereby significantly reducing false detections. Figure 3 (c) and Figure 3 (d) shows the detection results of the traditional target detection method and the method of the present invention in a dense near-shore scene. In this type of scene, ship targets are densely distributed, have large differences in scale, and are easily affected by shoreline structure and strong scattering background. The method of the present invention can effectively distinguish adjacent ship targets and still maintain good detection integrity for densely distributed small-scale ship targets, reducing target merging and missed detection.
[0038] The experimental results above demonstrate that the method of this invention, under conditions of limited manual annotation resources, can fully utilize massive amounts of unlabeled SAR images through a collaborative design of saliency pseudo-annotation, teacher-student network consistency semi-supervised learning, and uncertainty active sampling. It effectively reduces annotation costs while maintaining a high recall rate, significantly improving detection accuracy and verifying the effectiveness of the method. This method can be integrated into existing maritime surveillance and remote sensing analysis systems for ship target detection in ports, waterways, nearshore areas, and offshore areas. It is suitable for maritime supervision and fisheries law enforcement tasks under large-scale, long-term, and complex sea conditions, and has good engineering feasibility and application value.
[0039] Any aspects not covered in this invention are applicable to existing technologies.
Claims
1. A semi-supervised SAR ship target detection method based on pseudo-labeling and active learning, characterized in that, Includes the following steps: Step 1: Obtain a small number of manually labeled SAR images and a large number of unlabeled SAR images to construct labeled SAR image datasets and unlabeled SAR image datasets; Step 2: Randomly sample some unlabeled SAR images and perform saliency pseudo-labeling to obtain saliency pseudo-labeled SAR images, and add the saliency pseudo-labeled SAR images to the labeled SAR image dataset; For pixels in unlabeled SAR images, a multi-scale sliding window is used to obtain the neighborhood, the pixel amplitude within the sliding window is calculated, and the pixel amplitudes within the multi-scale sliding window are fused to obtain the brightness saliency features of the pixel. By iterating through all pixels, the brightness saliency features of the unlabeled SAR image are obtained; The gradient operator is used to calculate the mean gradient response of unlabeled SAR images under a multi-scale sliding window. The mean gradient response under the multi-scale sliding window is then fused to obtain the edge saliency features. Calculate the local texture statistics of unlabeled SAR images under a multi-scale sliding window, and fuse the local texture statistics under the multi-scale sliding window to obtain texture saliency features; After normalizing the saliency features, edge saliency features, and texture saliency features, a weighted fusion is performed to obtain the saliency map. The saliency graph is binarized, and then connected component analysis is performed to obtain multiple connected components. The area and aspect ratio of each connected component are calculated. Connected components whose area or aspect ratio is not within the preset range are identified as sparse connected components and removed, while those that are identified as dense connected components are retained. The DBSCAN clustering algorithm is used to cluster pixels within densely connected regions to generate multiple pixel clusters. The minimum bounding rectangle of each pixel cluster is used as the bounding box, and the saliency response intensity is used as the confidence level to obtain the saliency pseudo-labeled bounding box. All densely connected regions are traversed to obtain the saliency pseudo-labeled SAR image. Step 3: Construct a target detection model, perform consistent semi-supervised training on the target detection model based on a teacher-student network architecture, and use the trained target detection model for SAR ship target detection. The teacher network is pre-trained using an annotated SAR image dataset, the student network is initialized based on the parameters of the pre-trained teacher network, and the student network is then trained under supervision using the annotated SAR image dataset. Weak and strong enhancements are performed on the same unlabeled SAR image to obtain weakly enhanced and strongly enhanced unlabeled SAR images. The weakly enhanced unlabeled SAR image is passed to a pre-trained teacher network to obtain the predicted category, predicted bounding box, and predicted confidence. Predicted bounding boxes with a prediction confidence higher than a threshold are used as teacher pseudo-labels to constrain the unsupervised training of the student network. The strongly enhanced unlabeled SAR image is input into the student network for unsupervised training. During the semi-supervised training process, unlabeled SAR images were sampled from the unlabeled SAR image dataset through active learning of double uncertainty and manually labeled. The manually labeled SAR images were then added to the labeled SAR image dataset. The unlabeled SAR image dataset and the labeled SAR image dataset are updated, and the updated datasets are used in the subsequent training process. Dual uncertainty active learning sampling includes the following: The unlabeled SAR image is processed by the currently trained teacher network to generate candidate boxes, category prediction probabilities, and prediction confidence scores; the detection difficulty uncertainty score of the unlabeled SAR image is calculated according to equation (3): （3） In the formula, Indicates the first Uncertainty score for the detection difficulty of unlabeled SAR images. Indicates the first In the unlabeled SAR image, the first The predicted class probability of each candidate box; The information uncertainty score of the unlabeled SAR image is calculated according to equation (4): （4） In the formula, Indicates the first Uncertainty score of information content in unlabeled SAR images. Indicates the first In the unlabeled SAR image, the first The maximum prediction confidence of each candidate box; The uncertainty scores for detection difficulty and information content are normalized and then summed to obtain the comprehensive uncertainty score of the unlabeled SAR image. The unlabeled SAR image dataset is traversed to obtain the comprehensive uncertainty score of each unlabeled SAR image; several unlabeled SAR images with high comprehensive uncertainty scores are selected for manual annotation.
2. The semi-supervised SAR ship target detection method based on pseudo-labeling and active learning according to claim 1, characterized in that, During semi-supervised training, the total training loss of the student network is calculated using the following formula: （1） In the formula, This represents the total training loss of the student network; This indicates a loss of oversight. This represents the loss of unsupervised consistency. This represents the weighting coefficient.
3. The semi-supervised SAR ship target detection method based on pseudo-labeling and active learning according to claim 1, characterized in that, In the second step, the scales of the multi-scale sliding windows are 3×3, 7×7, and 11×11 pixels, respectively.
4. The semi-supervised SAR ship target detection method based on pseudo-labeling and active learning according to any one of claims 1 to 3, characterized in that, The target detection model used is the Faster R-CNN model.