An open set SAR target detection method combined with positioning quality estimation

Abstract

The application discloses an open set SAR target detection method combined with positioning quality estimation, comprising the following steps: acquiring a SAR data set and expanding the same to obtain a training data set; constructing an open set SAR target detector based on a single-stage target detection model, adding an unknown class discrimination branch to an existing classification head to determine whether a target belongs to a known class or an unknown class; meanwhile, adding a positioning quality estimation branch to an existing regression head to perform target positioning quality estimation regardless of the class; introducing unknown class probability learning loss and positioning quality estimation loss to train the open set SAR target detector, so as to use the trained open set SAR target detector to perform target detection. The method can not only detect and identify known classes, but also detect unknown class targets and judge them as unknown classes, thereby reducing the confusion between unknown class targets and known class targets and improving the performance of the SAR target detector in an open scene.

Description

Technical Field

[0001] This invention belongs to the field of SAR target detection technology, specifically relating to an open-set SAR target detection method that combines positioning quality estimation. Background Technology

[0002] Synthetic Aperture Radar (SAR) features all-weather, all-day operation and high resolution, enabling convenient acquisition of various surface information. As a fundamental problem in SAR image analysis and interpretation, SAR image target detection is crucial for real-time monitoring and information acquisition of various targets. With the rapid development of deep learning, increasingly more deep learning-based target detection methods are being applied to the field of SAR target detection, significantly improving its accuracy.

[0003] In recent years, due to the increasing depth of networks and the continuous improvement of computing power, deep learning-based object detection methods have begun to dominate. Currently, existing deep learning-based object detection algorithms often rely on labeled training data and are based on the closed-set assumption, which states that all target categories encountered during testing belong to categories already included in the training set. However, in real-world scenarios, new target classes often appear that are not present in or labeled in the training set. When dealing with open real-world scenes, object detectors often misclassify these new target classes as known classes, leading to a rapid decline in the detector's performance and making it difficult to meet practical requirements. SAR object detection also faces this problem, and it is even more severe because SAR images contain a large amount of background clutter, which can easily be misclassified as known classes. Therefore, SAR object detectors face even more severe performance degradation problems in open environments.

[0004] For target detection in open environments, in 2020, Akshay Dhamija et al. proposed a new target detection paradigm called open-set target detection. This paradigm requires the target detector to not only detect known target classes but also classify unlabeled new target classes as unknown. They also proposed an open-set target detection protocol to better evaluate the performance of the target detector under open-set conditions. In 2018, Miller et al. used dropout sampling to improve the robustness of the target detector under open-set conditions. In 2021, Joseph proposed an energy-based unknown class discriminator, which improves the performance of the target detector under open-set conditions by fitting the energy distributions of unknown and known classes. In 2022, Jiaming Han et al., based on the assumption that unknown targets are usually distributed in low-density regions of the latent space, proposed a method to achieve the discrimination of unknown targets by separating high / low-density regions in the latent space.

[0005] However, existing open-set target detection methods primarily target optical images. Due to the inherent complexity of SAR images and the significant influence of background clutter, a large amount of background clutter is often mistaken for unknown objects, leading to severe false alarms and further increasing the difficulty of identifying unknown targets in SAR target detection. Furthermore, compared to the abundant and diverse data in optical datasets, SAR datasets lack sufficient data and have fewer categories. During training, the model struggles to learn sufficiently generalized features, making it prone to overfitting to known targets and failing to detect unknown targets. In summary, existing open-set target detection methods cannot adequately address the SAR open-set target detection problem and fail to meet the requirements of SAR target detection in open scenes. Summary of the Invention

[0006] The purpose of this invention is to design an open-set SAR target detection method that, while maintaining detection accuracy for known target classes, enables the model to detect and identify unknown target classes, and minimizes clutter interference, thus meeting the requirements for SAR target detection in open scenes. Based on this, this invention provides an open-set SAR target detection method that incorporates positioning quality estimation. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] In a first aspect, this invention proposes an open-set SAR target detection method that combines positioning quality estimation, comprising:

[0008] The SAR dataset is acquired and expanded to obtain the training dataset; the training dataset includes multiple SAR training images with known class labels;

[0009] The open-set SAR target detector is constructed based on a single-stage target detection model. Specifically, it includes adding an unknown class discrimination branch to the classification head to determine whether the target belongs to a known class or an unknown class. At the same time, it adds a localization quality estimation branch to the localization head to perform target localization quality estimation that is independent of the class. This allows potential unknown class targets to be mined from the background of the training image based on the localization quality estimation results, and is used for training the unknown class discrimination branch.

[0010] The training dataset is input into the open-set SAR target detector. Based on the added unknown class discrimination branch and localization quality estimation branch, the unknown class probability learning loss and localization quality estimation loss are introduced to train the open-set SAR target detector so that the trained open-set SAR target detector can be used for target detection.

[0011] The beneficial effects of this invention are:

[0012] 1. The open-set SAR target detection method proposed in this invention combines positioning quality estimation. By improving the existing single-stage model, an unknown class discrimination branch is added to its classification head to determine whether the target belongs to a known class or an unknown class. At the same time, a positioning quality estimation branch is added to the positioning head to perform target positioning quality estimation independent of class, so as to mine potential unknown class targets from the background of the training image based on the positioning quality estimation results, and use it to train the unknown class discrimination branch, thus constructing an open-set SAR target detector. In addition, unknown class probability learning loss and positioning quality estimation loss are introduced, and the detector is trained using closed-set training. This allows the trained open-set SAR target detector to not only detect and identify known classes, but also detect unknown class targets and classify them as unknown classes, reducing the confusion between unknown and known class targets and improving the performance of the SAR target detector in open scenes.

[0013] 2. In response to the introduced unknown class discrimination branch, this invention designs an unknown class mining strategy based on hard-to-distinguish samples in known class positive samples and an unknown class mining strategy based on localization quality estimation. It can not only sample hard-to-distinguish samples from foreground targets to train the unknown class discrimination branch, but also mine potential unknown class targets from the background to train the unknown class discrimination branch, which significantly improves the detector's ability to detect and discriminate unknown class targets.

[0014] 3. When constructing the open-set SAR target detector, this invention adds a contrastive learning branch to the single-stage model classification head. Through contrastive learning, the distribution of known target classes in the latent space becomes more compact, thereby expanding the region of unknown target classes. This is more conducive to the discrimination of unknown target classes and further improves the model's ability to distinguish between known and unknown classes.

[0015] 4. The open-set SAR target detector constructed by the present invention using a single-stage RetinaNet model has an advantage in inference speed and is more practical compared to the existing two-stage model.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an open-set SAR target detection method combined with positioning quality estimation provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of the open-set SAR target detector built based on the single-stage RetinaNet model of this invention;

[0019] Figure 3 This is a schematic diagram of the unknown class probability learning target provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating an open-set SAR target detection method combining positioning quality estimation according to an embodiment of the present invention. The open-set SAR target detection method combining positioning quality estimation provided in this embodiment mainly includes the following steps:

[0023] Step 1: Obtain and expand the SAR dataset to obtain the training dataset; the training dataset includes multiple SAR training images with known class labels.

[0024] Specifically, images in the SAR training dataset can be obtained by augmenting the original SAR images. Common data augmentation methods mainly rely on external data or prior knowledge to generate new samples. Data augmentation allows the training dataset to include more comprehensive SAR training images, thereby improving the model's generalization ability, mitigating overfitting, and enhancing the model's robustness.

[0025] It should be noted that the embodiment constructs a closed-set training dataset here, and thus, each training sample in the training dataset, that is, the SAR training image, is labeled with a known type label.

[0026] Step 2: Construct an open-set SAR target detector based on a single-stage target detection model.

[0027] Since current open-set object detection algorithms often use Faster R-CNN as their base model, and Faster R-CNN is a two-stage object detector, its inference speed is slower than single-stage models such as RetinaNet and YOLOv3. Therefore, this embodiment chooses to build an open-set SAR object detector based on a single-stage object detection model. Compared to two-stage models, single-stage models have an advantage in inference speed and are more practical.

[0028] Optionally, as an implementation method, this embodiment preferably uses a single-stage RetinaNet model as the base model to construct an open-set SAR target detector.

[0029] Generally, a single-stage RetinaNet model mainly consists of three parts: the Backbone structure, the Neck structure, and the Head structure. The Head structure includes a classification head and a regression head. After the input image is fed into the model, it first passes through the Backbone structure for feature extraction at different levels, and then through the Neck structure for feature fusion at different levels, outputting feature maps of multiple scales. The feature maps of different scales are then processed through multiple convolutions in the Head structure, and then output for category and location information through the classification head and regression head, respectively.

[0030] This embodiment builds an open-set SAR target detector based on the single-stage target detector RetinaNet by improving its structure. Please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram of the structure of the open-set SAR target detector built based on the single-stage RetinaNet model of this invention.

[0031] First, to enable the model to discriminate unknown targets, this embodiment adds an unknown class discrimination branch to the single-stage RetinaNet classification head to determine whether the target belongs to a known class or an unknown class. Once the model is trained, the output value of this branch will be used during inference to determine whether the target belongs to a known class or an unknown class.

[0032] Optionally, the unknown class discrimination branch is derived from the penultimate feature layer of the single-stage RetinaNet model classification head. Its structure adopts a common convolutional structure, such as a 3*3 convolutional structure, and each detection box corresponds to one output.

[0033] Secondly, this invention is based on the assumption that potential unknown class targets will exist in the dataset during both the training and testing processes. Since these unknown class targets are unlabeled but require training of an unknown class discrimination branch, how to discover these potential unknown class targets and use them for training the unknown class discrimination branch is one of the key focuses of the proposed method. To enable the model to discover potential unknown class targets during training, this invention adds a branch to the RetineNet regression head, called the localization quality estimation branch. This branch is used for class-independent target localization quality estimation, allowing the model to discover potential unknown class targets from the background of the training images based on the output of this branch, and then use this information to train the unknown class discrimination branch.

[0034] Optionally, the localization quality estimation branch is derived from the penultimate feature layer of the regression head of the single-stage RetinaNet model, and like the unknown class discrimination branch, it adopts a common convolutional structure, such as a 3*3 convolutional structure, with each detection box corresponding to an output.

[0035] For further details, please continue to see Figure 2 The open-set SAR target detector constructed in this embodiment also includes a contrastive learning branch added to the classification head.

[0036] Specifically, given the consensus that known class targets are often distributed in high-density regions of the latent space, while unknown class targets are often distributed in low-density regions of the latent space, a contrastive learning branch is introduced to further improve the model's ability to distinguish between known and unknown classes. Through contrastive learning, the distribution of known class targets in the latent space becomes more compact, thereby expanding the low-density region, which will be beneficial for the discrimination of unknown class targets.

[0037] Optionally, the contrastive learning branch is derived from the penultimate feature layer of the single-stage RetinaNet model's classification head, specifically including two 3x3 convolutional layers and one L2 norm layer.

[0038] The contrastive learning branch added in this embodiment employs a contrastive clustering learning strategy, also known as a contrastive clustering learning head. The contrastive clustering learning head ultimately generates a 128-dimensional feature vector for each anchor (box), which is used for subsequent contrastive clustering learning.

[0039] Understandably, other contrastive learning strategies, such as instance-level contrastive learning, can also be used in practical applications.

[0040] Step 3: Input the training dataset into the open-set SAR target detector. Based on the added unknown class discrimination branch and localization quality estimation branch, introduce unknown class probability learning loss and localization quality estimation loss to train the open-set SAR target detector so that the trained open-set SAR target detector can be used for target detection.

[0041] Specifically, the training process for an open-set SAR target detector includes:

[0042] S1. Input the training image into the open-set SAR target detector to obtain the output of the localization quality estimation branch and the output of the unknown class discrimination branch.

[0043] It is understandable that, after the input training image is processed by the open-set SAR target detector, the output of the contrastive learning branch (contrastive clustering learning head), as well as the original classification head output and regression head output, will be obtained.

[0044] S2. Calculate the positioning quality estimation loss based on the output of the positioning quality estimation branch to train the positioning quality estimation branch; at the same time, calculate the unknown class probability learning loss based on the output of the unknown class discrimination branch to train the unknown class discrimination branch.

[0045] Since the open-set SAR target detector constructed in this invention requires the introduction of contrastive clustering loss during the training phase to train the contrastive learning branch.

[0046] The learning processes of the positioning quality estimation loss, unknown class probability learning loss, and contrastive clustering loss proposed in this invention will be described in detail below.

[0047] I. Positioning Quality Estimation Loss

[0048] In this embodiment, the positioning quality estimation loss is calculated based on the output of the positioning quality estimation branch in order to train the positioning quality estimation branch, specifically including:

[0049] Based on the output of the localization quality estimation branch, select N predicted boxes from all anchor boxes whose IOU (Intersection over Union) value with the ground truth (real label) is greater than a first preset threshold;

[0050] Randomly select a first preset number of prediction boxes from N prediction boxes, and use the IOU value between the selected prediction box and the ground truth as the label of the prediction box to train the localization quality estimation branch.

[0051] The L1 Loss function is used to calculate the positioning quality estimation loss. The calculation formula is as follows:

[0052]

[0053] In the formula, L OLN S represents the localization quality estimation loss, IOU represents the intersection-union ratio of the predicted bounding box and the ground truth bounding box, and S represents the localization quality estimation loss. i This represents the output of the localization quality estimation branch.

[0054] In this embodiment, the first preset threshold can be set to 0.3, and the first preset quantity can be set to N / 4.

[0055] II. Probabilistic Learning Loss for Unknown Classes

[0056] Since unknown class targets lack supervised information, it is necessary to study how to mine potential unknown class targets to train this branch. Based on this, this implementation adopts two methods for unknown class mining: the first is an unknown class mining strategy based on hard-to-distinguish samples in known positive samples, and the second is an unknown class mining strategy based on localization quality estimation. These two strategies are described in detail below.

[0057] Regarding the first strategy, since hard-to-distinguish samples among the known positive samples often lie on the decision boundary between the known and unknown classes, they are easily confused with the unknown class target. Furthermore, they can be considered to have features closer to the unknown class target. Therefore, selecting these hard-to-distinguish samples to train the unknown class discrimination branch is meaningful. Of course, for hard-to-distinguish samples, the model still needs to be able to classify them as their true class, rather than as the unknown class. Therefore, for these samples, it is hoped that the probability distribution vector output by the model for these samples after final training will be as follows... Figure 3 As shown, the probability of its true class is still the highest, but the probability of belonging to the unknown class is the second highest. This indicates that while maintaining the ability to accurately distinguish known class targets, the unknown class discrimination branch can indeed evaluate whether the target is an unknown class, thus achieving the goal of open set target detection.

[0058] Based on this, this embodiment proposes an unknown class mining strategy based on hard-to-distinguish samples in known positive samples, as follows:

[0059] During the training process of each epoch, based on the output of the unknown class discrimination branch, the IOU value between each anchor and the ground truth is calculated and compared with the second preset threshold to select positive and negative samples of the known class; wherein, the second preset threshold can be arbitrarily set according to actual needs.

[0060] Select hard-to-classify samples from the positive samples of the known classes, and assign a label of 1-P to each hard-to-classify sample. c This is used to train the unknown class discrimination branch; where hard-to-distinguish samples include samples located at the decision boundary between the known class and the unknown class; P c This represents the predicted probability of the ground truth category corresponding to the predicted bounding box of the difficult-to-distinguish sample.

[0061] Specifically, this embodiment can select difficult-to-distinguish samples from positive samples of known classes in the following manner:

[0062] For positive samples of a known class, calculate the cross-entropy according to the probability distribution vector output by its classifier head, then sort the cross-entropies, and finally select several samples with the largest cross-entropy (e.g., 30) as hard samples to train the unknown class discrimination branch.

[0063] The loss function for the unknown class mining strategy based on hard-to-distinguish samples in the known positive samples is calculated according to the following formula:

[0064]

[0065] In the formula, L UP1Let M represent the loss used to train the unknown class discriminant branch from the known positive samples. This represents the predicted probability of the ground truth category corresponding to the predicted bounding box of the j-th difficult-to-distinguish sample. This represents the output of the unknown class discrimination branch corresponding to the j-th difficult-to-distinguish sample.

[0066] Regarding the second strategy, this invention is based on the assumption that potential unknown class targets may exist in the image background during both the training and testing processes. Therefore, this invention proposes an unknown class mining strategy based on localization quality estimation, as follows:

[0067] First, it should be noted that the localization quality estimation branch introduced in this embodiment aims to train it to possess class-independent target discrimination capabilities. However, since the output of this branch is not accurate enough in the early stages of model training, this invention will use the output of this branch to find potential unknown class targets after the model has been trained for more than a second preset number of epochs (e.g., 100 epochs). At this point, two scenarios are considered:

[0068] (1) If there is ground truth in the output image of the current unknown class discrimination branch, select positive and negative samples from it; sort the negative samples according to the output of the localization quality estimation branch, select several samples with the largest localization quality estimation value (e.g., the first 10) as potential unknown class targets, and obtain the first sample set.

[0069] (2) If there is no ground truth in the output image of the current unknown class discrimination branch, select several samples with localization quality estimation scores greater than the third preset threshold and classification probability distribution vectors greater than the fourth preset threshold as potential unknown class targets to obtain the second sample set.

[0070] The reason for considering classification probability here is to reduce background clutter interference. Since all targets are negative samples at this time, if selection is based solely on the localization quality estimation score, a large number of background clutter will be selected. However, based on the assumption that there is feature similarity between unknown class targets and known class targets, targets with a classification probability vector greater than the fourth preset threshold are selected. The probability that these targets are background clutter will be very low, thereby reducing background clutter interference.

[0071] The third preset threshold can be set to 0.5, and the fourth preset threshold can be set to 0.4.

[0072] The first and second sample sets are combined into a potential unknown class sample set to train the unknown class discrimination branch. The loss function of the unknown class mining strategy based on localization quality estimation is then calculated according to the following formula:

[0073]

[0074] In the formula, L UP2 This represents the loss function used to train the unknown class discrimination branch using latent unknown class samples mined from the background, where λ represents the first hyperparameter, K represents the total number of samples in the latent unknown class sample set, and P... Uk This represents the output of the unknown class discrimination branch for the Kth potential unknown class sample.

[0075] Using the two strategies above, the probability learning loss for the unknown class can be obtained as follows:

[0076] L UP =L UP1 +L UP2 .

[0077] This invention addresses the introduced unknown class discrimination branch by designing an unknown class mining strategy based on hard-to-distinguish samples in known class positive samples and an unknown class mining strategy based on localization quality estimation. This strategy can not only sample hard-to-distinguish samples from foreground targets to train the unknown class discrimination branch, but also mine potential unknown class targets from the background to train the unknown class discrimination branch, significantly improving the detector's ability to detect and discriminate unknown class targets.

[0078] III. Comparison of Clustering Losses

[0079] Specifically, based on the consensus that known target classes tend to cluster in high-density regions of the latent space, while unknown target classes tend to be scattered in low-density regions of the latent space, if the distribution of known target classes in the latent space can be made more compact, then the low-density region can be expanded, thereby improving the model's ability to distinguish unknown target classes. Therefore, this paper proposes to introduce contrastive clustering loss to further improve the model's ability to distinguish unknown target classes.

[0080] The principle of contrastive clustering loss is to define a prototype belonging to each known class, and then bring the feature representation of the predicted box closer to the prototype of the same class and further away from the prototype of different classes. In this way, the feature representation of the predicted box in the latent space can be more compactly distributed, thereby expanding the low-density region to facilitate the discrimination of unknown class targets. At the same time, in order to enable the prototype to contain more diverse feature information, this invention introduces a contrastive clustering learner into the model, which consists of two components: a class-balanced memory bank, which stores a fixed number of feature vectors for each known class to improve the diversity of class features; and a contrastive clustering learning head, which consists of two 3*3 convolutional layers and an L2 Norm layer, which outputs a corresponding 128-dimensional feature vector for each anchor to represent the feature representation of the predicted box in the latent space.

[0081] Specifically, the learning process for the comparative clustering loss is as follows:

[0082] First, initialize a memory bank and create queues of the same capacity for each known class; for example, a queue with a capacity of 128 can be created.

[0083] Then, based on the output of the contrastive learning branch, in each training epoch, for each known class, a third preset number of feature vectors are sampled to update their respective queues; where the third preset number is less than the capacity of the queue; for example, when the capacity of the queue is 128, the third preset number can be set to 32.

[0084] Next, the feature vectors in the queue of each class are averaged to obtain the class prototype of that class, denoted as p. i .

[0085] Finally, based on the feature vectors in the class prototype and the queue, the contrastive clustering loss L is calculated. cc The calculation formula is:

[0086]

[0087]

[0088] In the formula, f c p represents the feature vector output by the contrastive learning branch in the queue, C represents the total number of feature vectors in the queue, and p i Let represent the class prototype, D() represent the Euclidean distance, Δ represent the second hyperparameter, and represents the minimum inter-class distance between different classes.

[0089] The learning process for the comparative clustering loss designed in this embodiment can also be shown in detail in Table 1 below.

[0090] Table 1 compares the learning process of clustering loss.

[0091]

[0092]

[0093] By constraining the loss function described above, the feature representations of proposals belonging to the same class can be brought closer to the class prototype of that class, while being moved further away from the class prototypes of other classes, thereby achieving the effect of making the intra-class compact and the inter-class separable.

[0094] In constructing an open-set SAR target detector, this invention adds a contrastive learning branch to the single-stage model classification head. Through contrastive learning, the distribution of known target classes in the latent space becomes more compact, thereby expanding the region of unknown target classes. This is more conducive to the discrimination of unknown target classes and further improves the model's ability to distinguish between known and unknown classes.

[0095] In addition, during the training phase, the classification head and regression head will be trained using the original classification loss and regression loss of the single-stage object detection model. For the specific implementation process, please refer to the existing related technologies. This embodiment will not be described in detail here.

[0096] S3. Based on the positioning quality estimation loss and the unknown class probability learning loss, and combined with the original classification loss and regression loss of the single-stage target detection model, the total loss of the open-set SAR target detector is obtained.

[0097] Specifically, based on the aforementioned positioning quality estimation loss L OLN Unknown class probability learning loss L UP And contrast clustering loss L CC And the original classification loss L cls and regression loss L reg The total loss L of the open-set SAR target detector can be obtained. all for:

[0098] L all =αL cls +βL reg +γL oln +δL CC +εL UP ;

[0099] In the formula, α, β, γ, δ, and ε represent the weight of each loss in the total loss.

[0100] S4. Update the network parameters of the open-set SAR target detector based on the total loss.

[0101] S5. Repeat steps S1-S4 until the open-set SAR target detector network converges, and a trained open-set SAR target detector is obtained.

[0102] In practical applications, the SAR image data to be detected is input into a trained open-set SAR target detector for processing to obtain the target detection result.

[0103] Understandably, after obtaining the trained open-set SAR target detector, the process also includes testing the open-set SAR target detector to evaluate its performance under different degrees of openness.

[0104] Specifically, this embodiment selects the SRSDD-v1.0 dataset for testing. This dataset contains a total of 6 classes: Container, Fishing, LawEnforce, cell-container, ore-oil, and Dredger. During testing, the dataset is divided into three types: the first is a closed set where all types are known; the second is a single-class open set where Dredger is the unknown class and the other 5 classes are known; and the third is a two-class open set where Dredger and cell-container are unknown classes and the other 4 classes are known. This allows for the evaluation of the performance of open set object detectors under different degrees of openness.

[0105] The open-set SAR target detector generated above was tested using the closed set, one type of open set, and two types of open sets, respectively. The results are shown in Tables 2, 3, and 4. Table 2 shows the test results using the closed set, as follows:

[0106] Table 2 Confusion Matrix of Closed Set Classes

[0107] Container 421 2 0 0 5 2 137 Fishing 12 30 1 0 0 1 39 LawEnforce 0 0 4 0 0 0 0 Cell-container 5 0 0 16 0 0 1 Ore-oil 9 0 0 1 22 0 4 Dredger 6 1 0 0 0 37 1 background 126 12 0 1 6 4 0

[0108] Table 3 shows the test results using a class of open sets, as follows:

[0109] Table 3. Confusion Matrix of a Class I Open Set Class

[0110]

[0111]

[0112] Table 4 shows the test results using two types of open sets, as follows:

[0113] Table 4. Confusion Matrix for Two Types of Open Sets

[0114]

[0115] As shown in Table 2-4, the performance of the open-set target detection model and the closed-set target detection model under one unknown class condition is very close, demonstrating the good performance of the proposed open-set target detection model. It can guarantee both the detection and recognition accuracy for known target classes and the detection and discrimination ability for unknown target classes. Secondly, the performance of the open-set target detection model under two unknown class conditions is worse than that under one unknown class condition, but it still maintains the detection and recognition accuracy for known classes and the detection and discrimination ability for unknown target classes. This shows that although the openness increases and the detection difficulty for unknown classes increases, the model proposed in this invention still possesses good performance.

[0116] Furthermore, this embodiment also compares the method of the present invention with two existing methods (Method 1 - Revisiting OpenWorld Object Detection; Method 2 - Random Boxes Are Open-world Object Detectors) in terms of unknown class detection rate, unknown class false alarm rate, and unknown class recognition rate. The results are shown in Table 5.

[0117] Table 5 shows the comparison results between the proposed method and existing methods on the SRSDDv1.0 dataset.

[0118]

[0119] As can be seen, compared with existing open-set target detection models for optical datasets, the open-set SAR target detector proposed in this invention has significant performance advantages.

[0120] This invention proposes an open-set SAR target detection method that combines positioning quality estimation. By improving existing single-stage models, an unknown class discrimination branch is added to the classification head to determine whether a target belongs to a known or unknown class. Simultaneously, a positioning quality estimation branch is added to the positioning head for class-independent target positioning quality estimation. This allows for the extraction of potential unknown class targets from the background of training images based on the positioning quality estimation results, which are then used to train the unknown class discrimination branch, thus constructing an open-set SAR target detector. Furthermore, an unknown class probability learning loss and a positioning quality estimation loss are introduced, and the detector is trained using closed-set training. This enables the trained open-set SAR target detector to not only detect and identify known classes but also detect and classify unknown class targets as unknown, reducing confusion between unknown and known class targets and improving the performance of the SAR target detector in open scenes.

[0121] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An open-set SAR target detection method combining positioning quality estimation, characterized in that, include: Obtain the SAR dataset and expand it to obtain the training dataset; The training dataset includes multiple SAR training images with known class labels; The open-set SAR target detector is constructed based on a single-stage target detection model. Specifically, it includes: adding an unknown class discrimination branch to the classification head to determine whether the target belongs to a known class or an unknown class; and adding a localization quality estimation branch to the regression head to perform target localization quality estimation that is independent of the class, so as to extract potential unknown class targets from the background of the training image based on the localization quality estimation results, and use them for training the unknown class discrimination branch. The training dataset is input into the open-set SAR target detector. Based on the added unknown class discrimination branch and localization quality estimation branch, unknown class probability learning loss and localization quality estimation loss are introduced to train the open-set SAR target detector so as to use the trained open-set SAR target detector for target detection. The training process of the open-set SAR target detector includes: The training image is input into the open-set SAR target detector to obtain the output of the localization quality estimation branch and the output of the unknown class discrimination branch; The positioning quality estimation loss is calculated based on the output of the positioning quality estimation branch in order to train the positioning quality estimation branch. The unknown class probability learning loss is calculated based on the output of the unknown class discrimination branch in order to train the unknown class discrimination branch; Based on the positioning quality estimation loss and the unknown class probability learning loss, and combined with the original classification loss and regression loss of the single-stage target detection model, the total loss of the open-set SAR target detector is obtained. The network parameters of the open-set SAR target detector are updated based on the total loss; Repeat the above steps until the open-set SAR target detector network converges, and a trained open-set SAR target detector is obtained. The positioning quality estimation loss is calculated based on the output of the positioning quality estimation branch in order to train the positioning quality estimation branch, specifically including: Based on the output of the localization quality estimation branch, anchors with an IOU value greater than a first preset threshold are selected from all anchors. N One prediction box; From the above N A first preset number of prediction boxes are randomly selected from the prediction boxes, and the distance between the selected prediction boxes and the ground truth is calculated. IOU The value is used as the label of the predicted bounding box to train the localization quality estimation branch; The L1 Loss function is used to calculate the positioning quality estimation loss, and the calculation formula is as follows: ； In the formula, This indicates the estimated loss in positioning quality. This represents the intersection-union ratio (IUU) between the predicted bounding boxes and the ground truth bounding boxes. This represents the output of the localization quality estimation branch; The unknown class probability learning loss is calculated based on the output of the unknown class discrimination branch to train the unknown class discrimination branch, including an unknown class mining strategy based on hard-to-distinguish samples in known positive samples, as follows: During the training process of each epoch, based on the output of the unknown class discrimination branch, the distance between each anchor and the ground truth is calculated. IOU The value is compared with a second preset threshold to select positive and negative samples of the known class; Select the hard-to-classify samples from the positive samples of the known classes, and set the label of each hard-to-classify sample to... The unknown class discrimination branch is used to train the unknown class discrimination branch; wherein the hard-to-distinguish samples include samples located at the decision boundary between the known class and the unknown class. This represents the predicted probability of the ground truth category corresponding to the predicted bounding box of a difficult-to-distinguish sample; The loss function for the unknown class mining strategy based on hard-to-distinguish samples in the known positive samples is calculated according to the following formula: ； In the formula, This represents the loss used to train the unknown class discrimination branch by utilizing hard-to-distinguish samples from the known positive class. M This represents the total number of difficult-to-distinguish samples. Indicates the first j The predicted probability of the ground truth category corresponding to each hard-to-distinguishable bounding box. Indicates the first j The output of the unknown class discrimination branch corresponding to each difficult sample.

2. The open-set SAR target detection method combining positioning quality estimation according to claim 1, characterized in that, Selecting difficult-to-distinguish samples from the known positive samples specifically includes: For positive samples of a known class, calculate the cross-entropy according to the probability distribution vector output by its classifier head, sort the cross-entropies, and finally select the samples with the largest cross-entropy as the hard-to-classify samples.

3. The open-set SAR target detection method combining positioning quality estimation according to claim 1, characterized in that, The unknown class probability learning loss is calculated based on the output of the unknown class discrimination branch to train the unknown class discrimination branch, and the method further includes: When the training epochs reach the second preset number, an unknown class mining strategy based on localization quality estimation is added, as follows: If ground truth exists in the output image of the current unknown class discrimination branch, positive and negative samples are selected from it; and the negative samples are sorted according to the output of the localization quality estimation branch, and the samples with the largest localization quality estimation values ​​are selected as potential unknown class targets to obtain the first sample set. If there is no ground truth in the output image of the current unknown class discrimination branch, select several samples with localization quality estimation scores greater than the third preset threshold and classification probability distribution vector sum greater than the fourth preset threshold as potential unknown class targets to obtain the second sample set; The first sample set and the second sample set are combined into a potential unknown class sample set to train the unknown class discrimination branch. The loss function of the unknown class mining strategy based on localization quality estimation is then calculated according to the following formula: ； In the formula, This represents the loss function used to train the unknown class discrimination branch by utilizing potentially unknown class samples mined from the background. Indicates the first hyperparameter. This represents the total number of samples in the potentially unknown class sample set. Indicates the first k The output of the unknown class discrimination branch for each potential unknown class sample.

4. The open-set SAR target detection method combining positioning quality estimation according to claim 1, characterized in that, The method for constructing an open-set SAR target detector based on a single-stage target detection model also includes: A contrastive learning branch is added to the classification head of the single-stage object detection model. This branch is used to make the distribution of known target classes in the latent space more compact through contrastive learning, thereby expanding the region of unknown target classes.

5. The open-set SAR target detection method combining positioning quality estimation according to claim 4, characterized in that, The contrastive learning branch employs a contrastive clustering learning strategy; correspondingly, during the training phase, a contrastive clustering loss is introduced to train the contrastive learning branch.

6. The open-set SAR target detection method combining positioning quality estimation according to claim 5, characterized in that, The learning process for the contrastive clustering loss includes: Initialize a memory bank and create queues of equal capacity for each known class; Based on the output of the contrastive learning branch, in each training epoch, for each known class, a third preset number of feature vectors are sampled to update their respective queues; wherein, the third preset number is less than the capacity of the queue. The class prototype of each class is obtained by averaging the feature vectors in the queue of each class. The contrastive clustering loss is calculated based on the class prototype and the feature vectors in the queue. The calculation formula is: ； ； In the formula, This represents the feature vector output by the contrastive learning branch in the queue. This represents the total number of feature vectors in the queue. Represents the class prototype. Represents Euclidean distance. This represents the second hyperparameter, which represents the minimum inter-class distance between different classes.

Images