A robust semi-supervised SAR ship detection method under annotation scale and quality constraints

Abstract

The present application relates to the technical field of image recognition, and more particularly to a robust semi-supervised SAR ship detection method under the constraints of annotation scale and quality. The method uses a multi-instance learning mechanism to correct inaccurate annotations containing noise through a label correction unit; then a teacher-student semi-supervised architecture is constructed to perform weak data enhancement and SAR-specific enhancement processing on unlabeled samples; during the training process, a decoupling training strategy is adopted to filter the pseudo-labels of the classification and regression tasks through different confidence thresholds; at the same time, a receptive field relaxation strategy is introduced in the label correction and semi-supervised learning unit, the scale of the region proposal is expanded through a relaxation factor, and multi-scale features are aggregated to enhance the resistance to background interference; finally, the teacher network is updated by exponential moving average. The present application effectively solves the problem of limited annotation scale and low quality of SAR images, and significantly improves the robustness and positioning accuracy of ship detection.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a robust semi-supervised SAR ship detection method under annotation size and quality constraints. Background Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave remote sensing imaging system, possesses unique advantages such as all-weather, all-time operation, strong penetration, and immunity to cloud and fog interference, playing an irreplaceable role in fields such as ocean-going ship monitoring, port traffic management, and maritime search and rescue. In recent years, deep convolutional neural networks have made significant progress in SAR ship target detection tasks. However, the training of these high-performance detection models often heavily relies on large-scale, high-quality manually labeled datasets.

[0003] In practical applications, obtaining ideal SAR training data faces significant challenges. Firstly, there are limitations in annotation cost and scale. Due to the distinct imaging mechanism of SAR images compared to optical images, their unique scattering characteristics and speckle noise make it difficult for non-experts to accurately identify ship targets. Accurately annotating bounding boxes on massive amounts of SAR images requires substantial human resources with specialized background knowledge, resulting in a often very limited number of labeled samples. Secondly, there are issues with annotation quality and noise interference. Limited by the low signal-to-noise ratio of radar images, blurred target edges, and multipath reflections, even professional annotators cannot guarantee that every bounding box perfectly matches the actual physical edges of the ship target. Real-world datasets often contain a large number of inaccurate annotations; such low-quality annotations directly lead to severe biases in supervised learning models, reducing positioning accuracy.

[0004] To alleviate the reliance on large-scale labeled data, semi-supervised target detection techniques have emerged. However, existing semi-supervised methods are mainly designed for high-quality labels, and have the following shortcomings when processing SAR data that simultaneously faces the dual constraints of limited scale and low quality: False label noise superposition: The existing framework cannot identify and correct the initial label noise in the original data, causing erroneous information to spread and amplify continuously in the iteration of the teacher-student architecture.

[0005] Poor environmental adaptability: SAR images have complex backgrounds, and strong scattering of land targets or interference from sea conditions can easily lead to false detections.

[0006] Imbalance between localization and classification: Existing methods often use a single confidence threshold when generating pseudo-labels, ignoring the differences in confidence requirements for SAR ship targets in classification and regression tasks. Summary of the Invention

[0007] To overcome the above shortcomings, this invention provides a robust semi-supervised SAR ship detection method under annotation scale and quality constraints. The aim is to design a robust semi-supervised detection scheme that can automatically correct initial annotation noise, efficiently utilize a large amount of unlabeled data, and optimize for SAR physical characteristics, so as to improve the utilization of SAR remote sensing images.

[0008] This invention provides the following technical solution: a robust semi-supervised SAR ship detection method under annotation size and quality constraints, comprising: S1. Obtain a SAR image dataset, which includes labeled samples and unlabeled samples containing inaccurate annotation information; S2. Input the labeled samples into the label correction unit, and use the multi-instance learning mechanism to correct the inaccurate initial annotations to obtain the corrected label samples. S3. Construct a teacher-student semi-supervised architecture, perform weak data augmentation and SAR specificity enhancement on unlabeled samples respectively, and input them into the teacher network and student network accordingly; S4. Use the teacher network to generate prediction results for unlabeled samples, and divide the prediction results into pseudo-labels corresponding to different tasks based on decoupled training. S5. In the label correction unit and the teacher network and student network, the region proposals generated by the region proposal network are scaled by a relaxation factor and multi-scale features are aggregated to enhance robustness to label noise and background interference. S6. Using the corrected label samples and the screened pseudo-labels, construct supervised loss and unsupervised loss respectively, jointly optimize the student network, and update the teacher network parameters through exponential moving average. S7. Use the trained model to extract ship target information from SAR images.

[0009] Preferably, the label correction unit is built based on Faster R-CNN and includes an instance generator, an instance classifier, and an instance selector, wherein the instance classifier and the instance selector share network parameters.

[0010] Preferably, the step of correcting inaccurate initial annotations using a multi-instance learning mechanism includes: Each inaccurately labeled target is considered a positive bag, and the background box is considered a negative bag; The positive package is expanded in multiple stages to obtain the expanded set of positive packages; The optimal instance is selected from the expanded positive package set by target-aware instance selection, and the optimal instance is used to guide the joint optimization of the instance generator, instance classifier and instance selector.

[0011] Preferably, the weak data augmentation includes at least one of random horizontal flipping and scaling of the unlabeled samples input to the teacher network; The SAR-specific enhancement includes at least one of speckle noise enhancement and image occlusion enhancement for unlabeled samples input to the student network; wherein the speckle noise enhancement is achieved by adding multiplicative noise that follows a gamma distribution to the image, and the image occlusion enhancement is achieved by randomly occluding the target region in the image with rectangular pixels.

[0012] Preferably, the step of dividing the prediction results into pseudo-labels corresponding to different tasks based on decoupled training includes: Pseudo-labels for classification tasks are selected from the prediction results based on a first confidence threshold. Pseudo-labels for the regression task are selected from the prediction results based on the second confidence threshold. The second confidence threshold is higher than the first confidence threshold.

[0013] Preferably, the scaling of the region proposals generated by the region proposal network is performed using a relaxation factor, including: The original region proposal is scaled up using the first relaxation factor and the second relaxation factor, respectively, to obtain two extended region proposals at different scales.

[0014] Preferably, the aggregated multi-scale features include: The original region proposal and the extended region proposal are input into the RoI alignment operator, and a feature map fused with multi-scale information is obtained through feature concatenation and convolution operations. The feature map is recalibrated along the channel dimension using a squeeze-excitation network.

[0015] Preferably, the unsupervised loss includes: We construct unsupervised classification loss using pseudo-labels from classification tasks, and unsupervised regression loss using pseudo-labels from regression tasks. The sum of the unsupervised classification loss and the unsupervised regression loss is taken as the unsupervised loss.

[0016] Preferably, updating the teacher network parameters via exponential moving average includes: After each training iteration of the student network, a weighted average of the student network parameters at the current moment and the teacher network parameters at the previous moment is calculated using a preset smoothing coefficient, and the result is used as the updated teacher network parameters at the current moment.

[0017] The present invention has the following beneficial effects: 1. This invention is the first to simultaneously solve the dual bottlenecks of limited sample size and low annotation quality in SAR ship detection. Through the LCRR-Det framework, multi-instance learning is used to effectively correct initial noisy labels, achieving efficient model training with limited and imprecise labeled data.

[0018] 2. The SAR-specific enhancement specially designed in this invention simulates typical interference in radar imaging, improving the model's generalization ability to complex sea conditions; in conjunction with the decoupled training strategy, it specifically optimizes the pseudo-label quality of classification and regression tasks, significantly enhancing the positioning accuracy of ship bounding boxes.

[0019] 3. The RFR strategy of this invention, through relaxation factors and multi-scale feature aggregation, endows the network with a broader context-aware perspective. This not only effectively suppresses complex background noise in SAR images, but also significantly improves the model's robustness to original label errors and semi-supervised pseudo-label noise. Attached Figure Description

[0020] Figure 1 This is a flowchart of a robust semi-supervised SAR ship detection method under annotation size and quality constraints proposed in this invention. Figure 2 This is an architecture diagram of the LCRR-Det framework proposed in this invention; Figure 3 This is a diagram of the tag correction unit framework in the LCRR-Det framework proposed in this invention; Figure 4 This is a diagram of the decoupled training framework in the LCRR-Det framework proposed in this invention; Figure 5 This is a diagram of the RFR strategy architecture in the LCRR-Det framework proposed in this invention; Figure 6 This is a visualization of the SAR ship detection performance of the LCRR-Det method proposed in this invention and its comparison method under weak marker noise mode; Figure 7 This is a visualization of the SAR ship detection performance of the LCRR-Det method proposed in this invention and its comparison method under strong labeled noise mode. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 In a first embodiment of the present invention, the present invention provides a robust semi-supervised SAR ship detection method under annotation size and quality constraints, such as... Figure 1 As shown, it includes the following steps: S1. Obtain a SAR image dataset, which includes labeled samples and unlabeled samples containing inaccurate annotation information; Specifically, this solution proposes an LCRR-Det framework for SAR ship target-level visual perception tasks, the structure of which is shown in the figure below. Figure 2 As shown. The training samples S of LCRR-Det include labeled samples. and unlabeled samples Two parts, of which and These represent the total number of labeled samples and the total number of unlabeled samples, respectively. (In the labeled samples...) middle, Represents an image. The bounding box represents the label information of an image. This represents the number of labeled samples. It's important to note that the labels here... It contains noise, meaning the bounding box cannot accurately enclose the target in the image. Unlabeled samples. The image is not included. The label information, and typically the number of images in unlabeled samples. Greater than the number of images in the labeled samples Therefore, this scheme takes into account the limitations of both tag size and quality, aiming to achieve target-level perception of ships in SAR images with only a small number of data tags, and is robust to inaccurate data labeling.

[0023] S2. Input the labeled samples into the label correction unit, and use the multi-instance learning mechanism to correct the inaccurate initial annotations to obtain the corrected label samples. Preferably, the label correction unit is built based on Faster R-CNN and includes an instance generator, an instance classifier, and an instance selector, wherein the instance classifier and the instance selector share network parameters.

[0024] Preferably, the step of correcting inaccurate initial annotations using a multi-instance learning mechanism includes: Each inaccurately labeled target is considered a positive bag, and the background box is considered a negative bag; The positive package is expanded in multiple stages to obtain the expanded set of positive packages; The optimal instance is selected from the expanded positive package set by target-aware instance selection, and the optimal instance is used to guide the joint optimization of the instance generator, instance classifier and instance selector.

[0025] Specifically, the label correction unit structure of the LCRR-Det framework is as follows: Figure 3 As shown, this unit is built on Faster R-CNN and includes an instance generator, an instance classifier, and an instance selector. The instance classifier and instance selector share network parameters. The regressor in the second stage of Faster R-CNN acts as the instance generator, used to generate candidate instances; both the instance classifier and instance selector are classifiers used to classify and select candidate instances.

[0026] The subsequent process of correcting inaccurate initial annotations using a multi-instance learning mechanism includes the following steps: First, object bags are constructed based on the output of the second stage of Faster R-CNN. Specifically, each inaccurately labeled object is treated as a positive bag, and the background bounding box is treated as a negative bag. Each positive bag contains multiple candidate instances derived from the original labeled bounding box and its extended region.

[0027] Secondly, the positive package is expanded in multiple stages to obtain an expanded set of positive packages. Specifically, the initial positive package is expanded using a multi-stage expansion method to generate a set of expanded positive packages, denoted as... ,in This represents the number of expanded positive packets. These expanded positive packets may contain high-quality instances that can be selected in subsequent steps to provide accurate monitoring signals.

[0028] Then, a target-aware instance selection mechanism is used to comprehensively consider the expanded positive packet set and the noisy ground truth labeled instances, selecting the optimal instance from among them. This selection mechanism can filter out the instance that best matches the real target from multiple candidate instances, serving as a supervision signal for subsequent training.

[0029] Finally, the selected optimal instance guides the joint optimization of the instance generator, instance classifier, and instance selector. Structurally, the instance classifier and instance selector are both classifiers and share parameters, while the regressor in the second stage of Faster R-CNN is considered an instance generator. The label correction unit described above enables the labeling of samples. Correcting inaccurate labels in the training data leads to more accurate training samples. In this study, the corrected sample is represented as... , where represents the label of the corrected sample.

[0030] S3. Construct a teacher-student semi-supervised architecture, perform weak data augmentation and SAR specificity enhancement on unlabeled samples respectively, and input them into the teacher network and student network accordingly; Preferably, the weak data augmentation includes at least one of random horizontal flipping and scaling of the unlabeled samples input to the teacher network; The SAR-specific enhancement includes at least one of speckle noise enhancement and image occlusion enhancement for unlabeled samples input to the student network; wherein the speckle noise enhancement is achieved by adding multiplicative noise that follows a gamma distribution to the image, and the image occlusion enhancement is achieved by randomly occluding the target region in the image with rectangular pixels.

[0031] Specifically, the label correction unit is designed to address the problem of inaccurate labels in labeled samples, and can overcome, to some extent, the annotation inaccuracies in SAR ship detection caused by crowdsourced annotation methods, the professional competence of annotators, imaging mechanisms, and environmental interference. Furthermore, the LCRR-Dot framework proposed in this scheme also includes a semi-supervised learning unit, consisting of supervised and unsupervised branches, which modify the corrected samples at a preset ratio. Images in unlabeled samples Data batches are generated for training. The supervised branch takes the corrected samples as input. Its training is inherited from the standard Faster R-CNN. The unsupervised branch takes unlabeled samples as input. First, by performing data augmentation processing separately, Get and .

[0032] For unlabeled samples input from the teacher network Weak data augmentation is employed. The purpose of weak data augmentation is to preserve the original structural information of the image, enabling the teacher network to generate relatively stable and reliable predictions. Weak data augmentation includes at least one of random horizontal flipping and scaling on the unlabeled samples input to the teacher network. Specifically, random horizontal flipping performs a horizontal mirror transformation of the image with a preset probability (e.g., 0.5); scaling scales the image with a random multiple (e.g., a random value between 0.5 and 1.5), allowing the network to better adapt to targets at different scales.

[0033] For unlabeled samples input into the student network SAR-specific enhancement is employed. This enhancement is applied only to the student network, not the teacher network, to establish an asymmetry between the two networks. SAR-specific enhancement includes at least one of speckle noise enhancement and image occlusion enhancement.

[0034] Speckle noise enhancement is used to simulate the coherence characteristics in the SAR imaging process. In this embodiment, the original image is first denoised to reduce the interference of the original noise on the enhancement process; then, multiplicative noise following a gamma distribution is added to the image to generate the enhanced image.

[0035] Image occlusion enhancement is used to simulate the occlusion effect caused by objects blocking radar signals. In this embodiment, a rectangular area in the image is randomly selected, and the pixel values ​​in the area are set to zero or a uniform gray value to simulate occlusion situations at different locations and ranges.

[0036] S4. Use the teacher network to generate prediction results for unlabeled samples, and divide the prediction results into pseudo-labels corresponding to different tasks based on decoupled training. Preferably, the step of dividing the prediction results into pseudo-labels corresponding to different tasks based on decoupled training includes: Pseudo-labels for classification tasks are selected from the prediction results based on a first confidence threshold. Pseudo-labels for the regression task are selected from the prediction results based on the second confidence threshold. The second confidence threshold is higher than the first confidence threshold.

[0037] Specifically, in semi-supervised learning units, the quality of pseudo-labels has a significant impact on the performance of the student network. Existing SSOD methods filter out low-quality bounding boxes from the teacher network's predictions by setting a fixed threshold, leaving relatively confident bounding boxes for the student network to train. However, this fixed threshold method treats classification and regression tasks equally, failing to consider the differences between these two tasks. According to relevant research, a lower threshold may be suitable for classification tasks, while regression tasks, due to their continuity, may require more refined supervision information. Furthermore, in SAR ship perception tasks where the size and quality of the labels being considered are limited, the corrected sample labels still have a certain offset compared to the finely labeled sample labels, posing a greater challenge to regression tasks than classification tasks. Therefore, this scheme employs decoupled training to consider the differences between classification and regression tasks, where the decoupled training framework is as follows: Figure 4 As shown, this scheme sets two different parameters. and These parameters are used to filter out pseudo-labels for classification and regression tasks, respectively. The parameter's role is to initially filter out most incorrect or low-quality bounding boxes from the teacher network's predictions, preventing them from generating inaccurate supervisory signals for the student network. Bounding boxes filtered by this parameter are used for the classification task. Bounding boxes with confidence scores higher than the parameter... The bounding box is considered to have higher quality and is therefore used to train regression tasks that require high accuracy of supervised information. In a preferred embodiment, the first confidence threshold... Set to 0.5, the second confidence threshold Set it to 0.8. This setting ensures that the recall rate for classification tasks is maintained while guaranteeing high-quality supervisory information for regression tasks.

[0038] S5. In the label correction unit and the teacher network and student network, the region proposals generated by the region proposal network are scaled by a relaxation factor and multi-scale features are aggregated to enhance robustness to label noise and background interference. Preferably, the scaling of the region proposals generated by the region proposal network is performed using a relaxation factor, including: The original region proposal is scaled up using the first relaxation factor and the second relaxation factor, respectively, to obtain two extended region proposals at different scales.

[0039] Preferably, the aggregated multi-scale features include: The original region proposal and the extended region proposal are input into the RoI alignment operator, and a feature map fused with multi-scale information is obtained through feature concatenation and convolution operations. The feature map is recalibrated along the channel dimension using a squeeze-excitation network.

[0040] Specifically, in the SAR ship detection task, which is the focus of this scheme and has limitations in both label accuracy and scale, the presence of label noise in the samples input to the label correction unit and the ambiguity of pseudo-labels generated by the teacher network pose challenges to the accurate perception of ships in SAR images. Furthermore, the background of ships in SAR images often contains unfavorable factors such as coastlines, speckle noise, sidelobes, and wakes, which also interfere with ship perception. To address this, this scheme proposes the Receptive Field Relaxation Strategy (RFR), which aims to fully utilize broader contextual information in the image to enhance the network's perceptual performance, thereby improving its adaptability to original data label noise, pseudo-label noise in the semi-supervised learning unit, and background interference, ultimately achieving accurate target perception.

[0041] like Figure 5 As shown, the RFR strategy in this scheme first scales the original region proposal R using relaxation factors E1 and E2, resulting in two additional scale region proposals, denoted as... and These two region proposals contain rich contextual information. Then, the original region proposal R is combined with the extended region proposal. and The input is fed into the RoI alignment operator, and through feature concatenation and convolution operations, a feature map aggregating multi-scale information is obtained. The above process can be represented as: ; in This indicates the RoI alignment operator. and These operations are feature concatenation and convolution. Furthermore, this scheme employs a squeeze-and-excitation network to aggregate multi-scale information from the feature map. Further processing is then performed. This network can compress channels and model the importance of each channel to achieve feature self-attention along the channel dimension, thus giving full attention to effective features. Overall, the RFR strategy described above, which includes scale expansion, feature aggregation, and squeeze-excitation networks, enables the network to better utilize contextual information, thereby enhancing the network's perceptual capabilities and improving its adaptability to label noise and background interference.

[0042] S6. Using the corrected label samples and the screened pseudo-labels, construct supervised loss and unsupervised loss respectively, jointly optimize the student network, and update the teacher network parameters through exponential moving average. Preferably, the unsupervised loss includes: We construct unsupervised classification loss using pseudo-labels from classification tasks, and unsupervised regression loss using pseudo-labels from regression tasks. The sum of the unsupervised classification loss and the unsupervised regression loss is taken as the unsupervised loss.

[0043] Preferably, updating the teacher network parameters via exponential moving average includes: After each training iteration of the student network, a weighted average of the student network parameters at the current moment and the teacher network parameters at the previous moment is calculated using a preset smoothing coefficient, and the result is used as the updated teacher network parameters at the current moment.

[0044] Specifically, the supervised branch uses the corrected labeled samples. The input is used, and its training method is consistent with the standard Faster R-CNN, preferably using the mean squared error loss function.

[0045] The unsupervised branch is constructed based on the teacher-student framework. The unsupervised loss consists of two parts: unsupervised classification loss and unsupervised regression loss, specifically represented as follows: ; in, This represents the image in the unlabeled sample. This indicates that the first confidence threshold has been met. Selected category pseudo-labels This indicates that the second confidence threshold has been met. The selected regression pseudo-labels, To represent the unsupervised classification loss, the cross-entropy loss function is preferred. This represents the unsupervised regression loss, and the smooth L1 loss function is preferred.

[0046] The total loss function is obtained by weighted summation of the supervised and unsupervised losses. : ; in, This indicates a loss of oversight. , Corresponding loss weights. In this implementation, the weight coefficients for both supervised and unsupervised losses are set to 1, meaning they are directly added together.

[0047] Specifically, the parameters of the teacher network do not directly participate in backpropagation; instead, they are obtained from the parameter updates of the student network through an exponential moving average (EMA). In each training iteration, the parameters of the teacher network are updated according to the following formula: ; in, Indicates the first The teacher network parameters are updated in rounds of iterations. Indicates the first The student network parameters are updated in rounds of iterations. The momentum coefficient is used to control the smoothness of the teacher network parameter updates. Through this method, the teacher network can maintain relatively stable predictive ability, providing high-quality pseudo-labels for unsupervised learning. In this embodiment, the momentum coefficient... The value range is from 0.99 to 0.999.

[0048] S7. Use the trained model to extract ship target information from SAR images.

[0049] Example 2 This embodiment verifies the effectiveness of the proposed method through experiments on the high-resolution SAR images dataset (HRSID). This dataset consists of 5604 SAR image samples from the ESA Sentinel-1 satellite and the German TerraSAR-X satellite. These images have an average size of 800 × 800 pixels, a spatial resolution ranging from 0.5 meters to 3 meters, and polarizations including HH, HV, and VV. In this embodiment, 35% of the images (1961 images) are randomly assigned to the test set, and 10% and 20% of the remaining 65% (3643 images) are randomly selected as training samples for LCRR-Det. Furthermore, considering that existing HRSIDs are finely annotated and contain few inaccurately labeled samples, this embodiment applies label perturbation to the training samples to simulate samples with inaccurate labels. By adjusting the perturbation amplitude, two modes are constructed: weakly labeled noise (20%) and strongly labeled noise (40%). Therefore, the label inaccuracy is greater in the strongly labeled noise pattern, making accurate detection more difficult.

[0050] This embodiment is built using PyTorch and employs an NVIDIA Tesla A100 GPU. Considering Faster R-CNN is the most representative object detection method, both the teacher and student networks in the label correction unit and the semi-supervised learning unit are designed based on Faster R-CNN, using a ResNet-50 pre-trained on ImageNet as the backbone network. During training, stochastic gradient descent (SGD) is used for optimization, with a learning rate of 2.5e-3, momentum of 0.9, and weight decay of 1e-4. SAR-specific augmentation in the semi-supervised learning unit is only applied to images input to the student model, while weak data augmentation is used on images input to the teacher network to increase the asymmetry between the two. Training of the supervised and unsupervised branches is performed simultaneously, with labeled and unlabeled samples randomly sampled at a 1:4 ratio to form a data batch for training. The label correction unit and the semi-supervised learning unit are trained step-by-step. Parameters used to calculate classification and regression losses during decoupled training are... and The values ​​were set to 0.5 and 0.8, respectively. The RFR strategy exists simultaneously in the teacher and student networks of both the label correction unit and the semi-supervised learning unit to fully adapt to the interference of noise in the annotation labels and pseudo-labels, as well as adverse factors in the SAR images. The scaling factor was set to empirical values ​​of 2 and 3. During the experiment, 10% and 20% of the samples were randomly selected from the training set and labeled with perturbations to obtain samples with inaccurate labels. The remaining images in the training set were unlabeled samples. Therefore, the training data consisted of a large number of unlabeled samples and a small number of samples with inaccurate labels, significantly reducing the dependence on the scale and quality of manual labeling training.

[0051] This approach uses the widely used MS-COCO dataset evaluation metrics to assess the detection performance of all methods. AP is the average of the predictions across 10 IoU thresholds linearly set between 0.50 and 0.95. and The average accuracy is calculated when the IoU threshold is set to 0.5 and 0.75, respectively. Evaluation metrics also include... , ,and ,in, Suitable for small ships with an area of ​​less than 32 square meters. Corresponding to medium-sized ships with an area between 32㎡ and 96㎡, Then focus on larger targets with an area greater than 96㎡.

[0052] This embodiment reports the performance of the proposed LCRR-Det in detecting ships in SAR images under both weakly labeled and strongly labeled noise modes. The supervised baseline in the comparative experiments was Faster R-CNN, trained using only labeled data. Several advanced and representative methods were used for comparison, including Unbiased Teacher, Soft Teacher, PseCo, Unbiased Teacher v2, and MixTeacher.

[0053] 1) Experimental Results in Weakly Labeled Noise Mode: Table 1 shows the quantitative experimental results of the proposed LCRR-Det and its comparative methods in SAR ship detection under weakly labeled noise mode. It can be seen that when the labeled sample ratio is 10%, the AP value of the supervised baseline is only 38.2, lower than LCRR-Det and other methods used for comparison. This is because the supervised baseline did not effectively utilize unlabeled samples during training, resulting in weak supervision information. The proposed LCRR-Det achieves an AP of 45.2 under a labeled sample ratio of 10%, outperforming other comparative methods. This indicates that LCRR-Det can achieve better detection performance for ships in SAR images. When the labeled sample ratio is 20%, the detection performance of each method is generally better than the corresponding method when the labeled sample ratio is 10%, which is consistent with the intuition that more labeled samples result in better detection performance. When the labeled sample ratio is 20%, LCRR-Det's AP value is 8.2 higher than the supervised baseline, and other indicators also outperform various methods used for comparison. This again reflects LCRR-Det's adaptability to conditions with limited labeled sample size and inaccurate labeling. It is worth mentioning that due to the relatively small number of large ships in HRSID, therefore... The values ​​are all relatively low. Overall, the AP value of this method is better than the comparison method in the weak label noise mode, indicating that it can adapt well to the limitations of labeled noise and lack of labeled samples.

[0054] Table 1. Quantitative experimental results of LCRR-Det and its comparative methods in weakly labeled noise mode.

[0055] This embodiment also visualizes the performance of LCRR-Det in SAR ship detection under weakly labeled noise mode. Due to space limitations, a basic supervisory baseline and a Mix Teacher with generally good quantitative experimental results were selected for qualitative comparison. Specific results are as follows: Figure 6As shown in the image, (a) represents the ground truth, and (b)-(d) represent the detection performance of the supervised baseline, MixTeacher, and LCRR-Det, respectively. Red ellipses in the image represent false negatives, green ellipses represent false positives, and purple ellipses represent overlapping bounding boxes. It can be seen that the supervised baseline exhibits a large number of false negatives and false positives, and overlapping bounding boxes are also observed in the complex scenes in the second column. This indicates that the supervised baseline performs poorly under the constraints of training sample size and quality. Compared to the supervised baseline, MixTeacher shows some improvement in false negatives and false positives. Furthermore, LCRR-Det shows no false negatives or false positives in the simple scenes of the first row, and only a small number of false negatives and false positives in the more complex scenes of the second to fourth rows. This good detection performance demonstrates that this scheme can effectively perceive ships in SAR images even with limited label size and quality.

[0056] 2) Experimental Results in Strong Label Noise Mode: Table 2 shows the quantitative experimental results of LCRR-Det and its comparative methods in strong label noise mode. It can be seen that the supervised baseline achieved APs of only 8.3 and 12.6, respectively, with labeled sample ratios of 10% and 20%. Although the detection performance of various comparative methods, including MixTeacher, was slightly better than the supervised baseline list, their overall performance was still poor, indicating that strong label noise can severely affect the detection performance of these methods. In contrast, LCRR-Det achieved detection performances of 33.6 and 37.2, respectively, with labeled sample ratios of 10% and 20%, far exceeding the baseline and all comparative methods, demonstrating its strong adaptability to samples containing strong label noise. Overall, this is attributed to the correction of inaccurate labels and a series of additional designs, resulting in performance far exceeding the supervised baseline and comparative methods in strong label noise mode, demonstrating excellent performance.

[0057] Table 2. Quantitative experimental results of LCRR-Det and its comparative methods under strongly labeled noise mode.

[0058] Figure 7The image demonstrates the target detection performance of ships in SAR images under weakly labeled noise mode. (a)-(d) visualize the detection results of the original image, Ground truth image, MixTeacher image, and LCRR-Det image, respectively. Red ellipses in the images represent missed detections, green ellipses represent false alarms, and purple ellipses represent overlapping detection boxes. It can be seen that in the first row of scenes, the supervised baseline misses most of the ships. While MixTeacher improves the missed detection, false alarms still occur. In contrast, LCRR-Det shows good detection performance, missing only a small number of tiny ships. In the dense scene shown in the second row, both the supervised baseline and MixTeacher show significant missed detections, while the proposed LCRR-Det detects the vast majority of targets, demonstrating good performance. Similarly, in the scene shown in the third row, the supervised baseline misses a large number of targets, and MixTeacher's detection performance is also unsatisfactory. In the scene shown in the fourth row, the supervised baseline misses four ships, and although MixTeacher only misses two, there is overlap in the predicted boxes. In contrast, LCRR-Det accurately detected all ships in the scene without any missed detections or false alarms. Overall, the proposed LCRR-Det is able to detect ships in SAR images better in strongly labeled noise mode.

[0059] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A robust semi-supervised SAR ship detection method under annotation size and quality constraints, characterized in that, include: S1. Obtain a SAR image dataset, which includes labeled samples and unlabeled samples containing inaccurate annotation information; S2. Input the labeled samples into the label correction unit, and use the multi-instance learning mechanism to correct the inaccurate initial annotations to obtain the corrected label samples. S3. Construct a teacher-student semi-supervised architecture, perform weak data augmentation and SAR specificity enhancement on unlabeled samples respectively, and input them into the teacher network and student network accordingly; S4. Use the teacher network to generate prediction results for unlabeled samples, and divide the prediction results into pseudo-labels corresponding to different tasks based on decoupled training. S5. In the label correction unit and the teacher network and student network, the region proposals generated by the region proposal network are scaled by a relaxation factor and multi-scale features are aggregated to enhance robustness to label noise and background interference. S6. Using the corrected label samples and the screened pseudo-labels, construct supervised loss and unsupervised loss respectively, jointly optimize the student network, and update the teacher network parameters through exponential moving average. S7. Use the trained model to extract ship target information from SAR images.

2. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 1, characterized in that, The label correction unit is built on Faster R-CNN and includes an instance generator, an instance classifier, and an instance selector, with the instance classifier and instance selector sharing network parameters.

3. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 2, characterized in that, The method of using a multi-instance learning mechanism to correct inaccurate initial annotations includes: Each inaccurately labeled target is considered a positive bag, and the background box is considered a negative bag; The positive package is expanded in multiple stages to obtain the expanded set of positive packages; The optimal instance is selected from the expanded positive package set by target-aware instance selection, and the optimal instance is used to guide the joint optimization of the instance generator, instance classifier and instance selector.

4. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 1, characterized in that, The weak data augmentation includes at least one of random horizontal flipping and scaling of unlabeled samples input to the teacher network; The SAR-specific enhancement includes at least one of speckle noise enhancement and image occlusion enhancement for unlabeled samples input to the student network; wherein the speckle noise enhancement is achieved by adding multiplicative noise that follows a gamma distribution to the image, and the image occlusion enhancement is achieved by randomly occluding the target region in the image with rectangular pixels.

5. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 1, characterized in that, The method of dividing the prediction results into pseudo-labels corresponding to different tasks based on decoupled training includes: Pseudo-labels for classification tasks are selected from the prediction results based on a first confidence threshold. Pseudo-labels for the regression task are selected from the prediction results based on the second confidence threshold. The second confidence threshold is higher than the first confidence threshold.

6. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 1, characterized in that, The scaling of the region proposals generated by the region proposal network is performed using a relaxation factor, including: The original region proposal is scaled up using the first relaxation factor and the second relaxation factor, respectively, to obtain two extended region proposals at different scales.

7. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 6, characterized in that, The aggregated multi-scale features include: The original region proposal and the extended region proposal are input into the RoI alignment operator, and a feature map fused with multi-scale information is obtained through feature concatenation and convolution operations. The feature map is recalibrated along the channel dimension using a squeeze-excitation network.

8. The robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 1, characterized in that, The unsupervised loss includes: We construct unsupervised classification loss using pseudo-labels from classification tasks, and unsupervised regression loss using pseudo-labels from regression tasks. The sum of the unsupervised classification loss and the unsupervised regression loss is taken as the unsupervised loss.

9. A robust semi-supervised SAR ship detection method under annotation size and quality constraints according to claim 1, characterized in that, The method of updating teacher network parameters through exponential moving average includes: After each training iteration of the student network, a weighted average of the student network parameters at the current moment and the teacher network parameters at the previous moment is calculated using a preset smoothing coefficient, and the result is used as the updated teacher network parameters at the current moment.

Images