A high-precision fine-grained SAR target detection method

Abstract

The application discloses a high-precision fine-grained SAR target detection method, which comprises the following steps: inputting a to-be-detected SAR image into a trained SAR image fine-grained detection model, and outputting the coordinate positions and the categories of all targets of interest in the to-be-detected SAR image; the training method of the model comprises the following steps: constructing a global instance set; taking a training data set as input; extracting the coordinate positions of the targets of interest in the training set by the model, and identifying the categories of the extracted targets of interest; combining the extracted and identified contents into a model collection data set; randomly extracting instances from the global instance set to form a sampling data set; respectively coding the targets of interest in the model collection data set and the sampling data set, and then performing global instance comparison; optimizing the model according to the comparison result until the final SAR image fine-grained detection model is obtained; and the method can reduce the sample quantity in the comparison learning method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and pattern recognition, and particularly relates to a high-precision fine-grained SAR target detection method. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is an active earth observation system, which can realize all-weather and all-time observation of the earth, and has unique advantages in disaster monitoring, environmental monitoring, ocean monitoring, mapping and military applications, and can play a role that other remote sensing methods cannot play. The target detection task is to give the coordinate positions of all targets of interest in the image and identify the specific category to which the target belongs, which is one of the important basic tasks of SAR image processing. Further, fine-grained detection requires the model to accurately identify similar but different categories of targets in the SAR image. Due to the lack of detailed information of the target in the SAR image, the inter-class difference of different targets is small, and how to more accurately perform fine-grained detection has both theoretical research value and practical application value.

[0003] Most of the current mainstream advanced target detection algorithms pay more attention to how to accurately locate the target, and cannot effectively handle the fine-grained classification and recognition problem in detection, which may result in accurate positioning but incorrect classification. Moreover, compared with optical images, SAR images lack detailed texture information due to the particularity of their imaging method, and are easily disturbed by noise and redundant background, which makes it more difficult for general detection algorithms to distinguish similar but different targets.

[0004] Contrastive learning has the effect of increasing the inter-class target difference and improving the intra-class target similarity, and is an effective method to enhance fine-grained classification performance. However, most of the current contrastive learning techniques are developed based on image-level classification tasks, and are concentrated in the field of unsupervised pre-training, and their application in supervised learning, especially in target detection tasks, is less. At the same time, some successful contrastive learning methods need to extract a large number of positive and negative sample pairs for contrastive training, and the wide characteristics of SAR images and the sparsity of aircraft target distribution make it impossible to extract a large number of samples at one time during the training of the detection model. In addition, the key of contrastive learning is to narrow the distance of feature vectors of the same category in the projection space, while making the features of different categories have sufficient difference. However, considering the demand of the positioning task for the features, too much pursuit of the difference of the features may damage the accuracy of the positioning.

[0005] Therefore, to realize accurate SAR fine-grained target detection, a contrastive learning method at the instance level is needed, which does not need to extract a large number of samples at one time during the training step, and at the same time, the influence on the accuracy of the positioning task needs to be modified, which also becomes a key problem of the current research. SUMMARY

[0006] In view of the above problems, the present invention provides a high-precision fine-grained SAR target detection method that at least solves some of the above technical problems, thereby reducing the number of samples in the contrastive learning method.

[0007] This invention provides a high-precision, fine-grained SAR target detection method, comprising:

[0008] Acquire the SAR image to be detected, and input the SAR image to be detected into the trained SAR image fine-grained detection model, and output the coordinate positions and categories of all targets of interest in the SAR image to be detected;

[0009] The training method for the SAR image fine-grained detection model is as follows:

[0010] S1. Construct a global instance set based on the training dataset;

[0011] S2. Construct a fine-grained SAR image detection model framework; input the training dataset into the fine-grained SAR image detection model framework; the fine-grained SAR image detection model framework extracts the coordinate positions of targets of interest in the training set and identifies the category of the extracted targets of interest; combine the extracted and identified content into a model acquisition dataset;

[0012] S3. Randomly select instances from the global instance set to form a sampling dataset;

[0013] S4. After encoding the target of interest in the model's collected dataset and the sampled dataset respectively, perform a global instance comparison;

[0014] S5. Optimize the SAR image fine-grained detection model framework based on the comparison results until the final SAR image fine-grained detection model is obtained.

[0015] Furthermore, S1 specifically includes:

[0016] Acquire SAR image training samples; the SAR image training samples include multiple target regions;

[0017] The training dataset consists of SAR image training samples, as well as the coordinates and categories of all known targets of interest within all target regions.

[0018] All target regions in the training dataset are cropped and scaled to form a global instance set.

[0019] Further, in the S2, the coordinate positions of the targets of interest in the training set are extracted by the SAR image fine-grained detection model framework, and the categories to which the extracted targets of interest belong are identified, specifically comprising:

[0020] The input SAR image training sample is subjected to feature map extraction by the backbone convolutional network in the SAR image fine-grained detection model framework, and a plurality of feature maps of different sizes are obtained by calculating a feature pyramid;

[0021] Each of the feature maps is sent to a detection head, and the coordinate positions of the targets of interest in each of the feature maps are extracted, and the categories to which the extracted targets of interest belong are identified.

[0022] Further, for each of the feature maps, the preparation work before global instance comparison includes:

[0023] According to the categories to which the input targets of interest belong, the feature map is cropped into a plurality of regions, denoted as a first feature region set;

[0024] Random sampling is performed on the target regions in the global instance set, and the backbone convolutional network is used to extract features of the sampled target regions, denoted as a second feature region set;

[0025] After the targets of interest in the first feature region set and the second feature region set are encoded respectively, global instance comparison is performed.

[0026] Further, an encoder composed of an average pooling and a fully connected layer is used to encode the targets of interest using tanh as a nonlinear activation function.

[0027] Further, the instance-level comparison loss used in the global instance comparison is defined as follows:

[0028]

[0029]

[0030] Wherein, L CTLi represents the comparison loss of each feature map; the numerator represents the sum of the intra-class feature similarities; the denominator represents the sum of the similarities of all features; f i and f j represent two different feature vectors; c i and c j represent the category labels of the corresponding vectors; i and j represent the traversal indication subscripts; cos i ,f jrepresents the cosine similarity between two features; tau represents the temperature parameter; M represents the number of target regions in the first feature region set; N represents the number of target regions in the second feature region set; c i represents the known class label corresponding to each feature; L CTL represents the contrast loss of the overall after averaging.

[0031] Further, the detection head uses an anchor-free structure of full convolution;

[0032] The detection head is used for predicting the position and class confidence of each target of interest on different scale feature maps.

[0033] Further, the detection head adopts a focus loss based on positioning quality perception as a classification loss;

[0034] The focus loss based on positioning quality perception is represented as:

[0035]

[0036] Wherein, q represents the intersection over union of the predicted bounding box and the real bounding box at the point; p represents the classification confidence of the current sample obtained by the SAR image fine-grained detection model; y=1 / 0 represents that the point belongs to or does not belong to the category; when y=0, the corresponding q will also be set to 0; t, beta, gamma are all hyperparameters for controlling the proportion of each loss.

[0037] Further, the SAR image fine-grained detection model adopts a two-stage positioning method;

[0038] In the first stage, the SAR image fine-grained detection model predicts the vector of all positive sample points, which respectively represents the rough distance prediction of the sample point to the four boundaries of the real bounding box.

[0039] In the second stage, based on the prediction result of the first stage, a boundary perception bounding box fine optimization module is used to adjust the bounding box.

[0040] Further, the overall loss function of the SAR image fine-grained detection model is composed of three parts: the classification loss, the positioning loss and the target positioning loss function.

[0041] Compared with the prior art, the high-precision fine-grained SAR target detection method has the following beneficial effects:

[0042] Firstly, the supervised contrast learning method is applied to the detection model, and a global instance sampling is proposed to increase the sample size of the contrast learning.

[0043] In the training process, the model will calculate the contrast loss of the target in the image and the sampled target together to enhance the representation learning ability.

[0044] To enhance the positioning ability of the model, the application designs an edge-aware fine bounding box optimization module. By extracting the boundary features above, below, left and right of the target respectively, the model can realize more accurate positioning.

[0045] The application proposes a positioning quality-aware focal loss, which gives greater weight to samples with accurate positioning but classification errors, so that the model pays more attention to fine classification problems under high-quality positioning.

[0046] Other features and advantages of the application will be described in the following description, and some will become apparent from the description, or will be understood by those skilled in the art. The purpose and other advantages of the application can be achieved and obtained by the structure specifically pointed out in the written description, claims and drawings.

[0047] The technical solutions of the application will be described in detail below with the help of the drawings and examples. DETAILED DESCRIPTION

[0048] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0049] Figure 1 The SAR image fine-grained detection model framework based on global instance contrast provided for the embodiment of the application is shown in the figure.

[0050] Figure 2 The edge-aware fine bounding box optimization provided for the embodiment of the application is shown in the figure.

[0051] Figure 3 The effect comparison diagram of global instance contrast on fine-grained detection performance provided for the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0052] The exemplary embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0053] The embodiment of the application provides a high-precision SAR image fine-grained target detection method, which can provide theoretical and technical support for the research of the fine-grained target detection field in the academic field, and can also be applied to the processing of SAR images and the extraction of interested targets in practice; the method specifically comprises the following steps:

[0054] Obtain a SAR image to be detected, and input the SAR image to be detected into a trained SAR image fine-grained detection model, to output the coordinate positions and the categories of all interested targets in the SAR image to be detected;

[0055] The training method of the SAR image fine-grained detection model is as follows:

[0056] S1, obtain a SAR image training sample; the SAR image training sample comprises a plurality of target regions; the SAR image training sample and the coordinate positions and the categories of all interested targets in all target regions are taken as a training data set; all target regions in the training data set are cropped and scaled to form a global instance set;

[0057] S2, construct a SAR image fine-grained detection model framework; input the training data set into the SAR image fine-grained detection model framework; the SAR image fine-grained detection model framework extracts the coordinate positions of the interested targets in the training data set and identifies the categories of the extracted interested targets; and the extracted and identified contents are combined into a model collection data set;

[0058] S3, randomly extract instances from the global instance set to form a sampling data set;

[0059] S4, after the interested targets in the model collection data set and the sampling data set are encoded, respectively, the global instances are compared;

[0060] S5, according to the comparison result, the SAR image fine-grained detection model framework is optimized until the final SAR image fine-grained detection model is obtained.

[0061] The above S1, S2, S3, S4 and S5 are only for understanding and do not limit the specific steps of the training method.

[0062] Next, the training method of the SAR image fine-grained detection model is specifically described.

[0063] (1) Task definition

[0064] The task definition of the SAR image fine-grained detection model is as follows: given a SAR image I, the trained model can infer the coordinate positions and the categories of all detected targets on the image:

[0065]

[0066] wherein model(I) represents a detection model; bbox i represents the i-th target of interest; respectively represent the position of the upper left corner and the lower right corner of the bounding box; c i represents the category to which the target of interest belongs; i represents the target of interest; and M represents the number of targets of interest.

[0067] (2) Overall architecture of the SAR image fine-grained detection model

[0068] The input of the model during training is a SAR image I and its corresponding target position and category label to be detected, which is represented as a set B:

[0069]

[0070] For the input SAR image, the model uses a backbone convolutional network G(·) to extract feature maps, and obtains four feature maps P2, P3, P4 and P5 of different sizes by calculating a feature pyramid. Subsequently, the feature maps of different sizes are sent to a detection head to extract the coordinate position of the target of interest in each feature map and identify the category to which the extracted target of interest belongs.

[0071] Taking the feature map P2 as an example, the target region to be detected in P2 is extracted, and random sampling is performed from the constructed global instance set, and the extracted target and the sampled instance are encoded and compared.

[0072] (3) Global instance contrast (GIC)

[0073] Before the model is trained, all target regions in the training data set are cropped and scaled to form a global instance set (such as Figure 1 the upper right corner). During training, first, according to the input label (i.e., the category to which the target of interest belongs), all target regions in the corresponding region P2 are cropped, denoted as F1, F2,..., F M . Secondly, N instances are randomly extracted from the global instance set, and features are extracted using the same G(·) to obtain F M+1 ,F M+2 ,...,F M+N .

[0074] In the embodiment of the present application, when encoding, an encoder composed of an average pooling layer and a fully connected layer is used to encode the target of interest using tanh as a nonlinear activation function:

[0075] f i i ))), i = 1, 2,..., M + N (1)

[0076] In the above formula, FC(·) represents a fully connected layer; avgpool(·) represents an average pooling calculation. For a target F i , f i is a 256-dimensional feature vector obtained after the fully connected layer, average pooling and tanh calculation. In addition, c i represents the known class label corresponding to each feature. Finally, the set F represents all target features and their class information used for supervised representation learning:

[0077] F = { <f1, c1>, <f2, c2>,..., <f M+N , c M+N >} (2)

[0078] In formula (2), <f1, c1>, <f2, c2>,..., <f M+N , c M+N > represents the feature vector and the class label of each target.

[0079] In the embodiments of the present application, the instance-level contrast loss used in global instance contrast is defined as follows:

[0080]

[0081]

[0082] Where L CTLi represents the contrast loss of each feature map; the numerator represents the sum of intra-class feature similarities; the denominator represents the sum of all feature similarities; f i , f j represents two different feature vectors, c i , c j represents the class label of the corresponding vector, i, j represents the traversal index; cos <f i , f j > represents the cosine similarity between two features; τ represents a temperature parameter; M represents the number of target regions in the first feature region set; N represents the number of target regions in the second feature region set; c i represents the known class label corresponding to each feature; for each f i , the model regards the samples of different classes as negative samples. L CTL ​This represents the overall contrastive loss obtained after averaging. Minimizing this contrastive loss increases the similarity within the same class at the instance level, while also increasing the difference in features between different classes. Class-specific features help the model achieve higher accuracy in downstream classification tasks.

[0083] (4) Edge perception fine optimization

[0084] In this embodiment, the detection head uses a fully convolutional anchor-free structure. The detector predicts the target's location and class confidence at each point on feature maps at different scales. For any real target (x1, y1, x2, y2, c) in the bounding box, treat all feature points falling within this bounding box as positive sample points and assign the true label c to the classification. T Let c be the value of the feature point. If a feature point is not within the bounding box of any ground truth bounding box, it will be considered a negative sample point and c will be set accordingly. T =0. For classification tasks, the model generates a confidence vector of CD for each sample point, where C is the total number of classes.

[0085] To enhance the detector's localization capability, in this embodiment of the invention, the SAR image fine-grained detection model employs a two-stage localization method; in the first stage, the SAR image fine-grained detection model predicts the vectors (l) for all positive sample points. 0 ,r 0 ,t 0 ,b 0 ), representing the coarse distance predictions from the sample point to the left, right, top, and bottom boundaries of the true bounding box, respectively; in the second stage, based on the prediction results of the first stage, the bounding box fine-tuning module is used to adjust the bounding box. For example... Figure 2 As shown, the four-branch separable convolution is used to extract features from the four boundary regions of the thick bounding box and predict the correction values ​​of the bounding box. Finally, the model computes a correction vector (Δl, Δt, Δr, Δb) for each point on the feature map, where the four parameters represent the correction coefficients of the original predicted values ​​of the four bounding box distances relative to the true values. The two-stage bounding box prediction result is given by the following equation:

[0086]

[0087] In equation (5), l, r, t, and b are the final predicted distances between the positive sample point and the four boundaries of the real bounding box, respectively.

[0088] For a positive sample point (x, y) on the feature map, its coordinates mapped to the input image are represented as follows: Where s is the scaling ratio of the feature map relative to the original image. is the coordinate representation on the input image. For the real bounding box that this sample belongs to, the true value of the two-stage prediction is: * = (xl,yl,x2,y2), the true value of the two-stage prediction corresponds to:

[0089]

[0090] (xl,yl,x2,y2) in formula (6) is the left upper corner and right lower corner coordinates of the real bounding box, is the coordinate representation on the input image of the sample point, (l * ,r * ,t * ,b * ) is the true value that the two-stage model needs to predict. The coarse bounding box and the final bounding box of the two-stage prediction are respectively bbox c = (l 0 ,r 0 ,t 0 ,b 0 ), bbox f = (l,r,t,b), this method uses GIOU loss to calculate the loss of the positioning task:

[0091] L reg = L GIOU (bbox c ,bbox * ) + L GIOU (bbox f ,bbox * ) (7)

[0092] In formula (7), bbox * ,bbox c ,bbox f are the true value to be predicted, the coarse bounding box predicted by the one-stage and the fine bounding box obtained by the two-stage, respectively, L GIOU is the GIOU loss function between the two bounding boxes, and L reg is the total loss of the positioning task.

[0093] (5) Positioning quality-aware focal loss

[0094] The mainstream one-stage detector currently mostly uses focal loss as the classification loss. However, focal loss does not consider the influence of positioning accuracy when calculating. In order to avoid the situation that the bounding box prediction is accurate but the classification is wrong, the classification loss should pay more attention to the prediction bounding box with high positioning accuracy, and improve the classification accuracy of high-quality bounding box. Therefore, the intersection over union (IOU) of the predicted bounding box and the true value is introduced, and an IOU-aware focal loss (IAFL) is proposed, which is defined as follows:

[0095]

[0096] For positive sample points, q represents the IOU of the predicted bounding box and the true bounding box at the point; p is the classification confidence of the current sample obtained by the model. The detector predicts a length C vector (p1, p2,..., p C ) for each point, which represents the probability of belonging to each class respectively. The model calculates the binary classification error for each class, and y=1 / 0 represents that the point belongs to / not belong to the class. When y=0, the corresponding q will also be set to 0. t, β, β in equation (8) are all hyperparameters for controlling the proportion of each item of the loss; in the embodiment of the present application, t=1, γ=1.5, and γ=1.

[0097] IAFL adds IOU weight q and bounding box quality coefficient (t+max(q-p,0)) β to the focal loss. The IOU weight q can guide the detector to pay attention to the sample with high positioning quality, and (t+max(q-p,0)) β will increase the value of the loss function when the positioning accuracy is high but the classification result is poor, and t is a parameter added to prevent this item from being 0.

[0098] (6) Overall loss function

[0099] The loss function of the model is composed of three parts: classification loss, positioning loss and target positioning loss function, and the overall loss function is as follows:

[0100]

[0101] Where K is the total number of sample points, P is a set of all positive sample points, i is the index of traversal, bbox c (i) ,bbox f (i) ,bbox *(i)are the coarse bounding box prediction, the fine bounding box prediction and the ground truth of the regression stage corresponding to the i-th sample point respectively; L reg is the regression loss defined in equation (7); IAFL is the localization quality-aware focal loss defined in equation (8), p i , i , i are the current sample classification confidence, the intersection over union of the current predicted bounding box and the corresponding ground truth, and the classification label of the current sample respectively; L CTL is the global contrastive learning loss defined in equation (4); L is the total loss obtained in a single training process of the model; equation (9) indicates that the localization loss is only calculated for positive sample points, while negative sample points only participate in the calculation of the classification loss.

[0102] Next, the above content will be specifically described through a specific embodiment.

[0103] SAR image fine-grained aircraft target detection

[0104] The SAR image data obtained by the GaoFen-3 remote sensing satellite is used for fine-grained aircraft target detection training, and compared with other target detection methods, and the results are shown in Table 1:

[0105] Table 1 Precision comparison of the method and other detection models in SAR image fine-grained aircraft detection

[0106]

[0107]

[0108] As shown in Table 1, the experimental results show that the high-precision fine-grained SAR target detection method provided by the application can well solve the SAR image fine-grained recognition problem, and almost surpasses the existing detection method on different backbone networks. The best model using ResneXt as the backbone network achieves 74.7 AP, 98 AP50 and 84.7 AP75, reaching the SOTA result of the SAR fine-grained aircraft detection task at the present stage. The experimental results fully verify the effectiveness of the method. It is worth mentioning that the AP75 of the model on ResNet-50 and ResNeXt-101 has a performance improvement of 3.0 / 0.6 compared with the suboptimal detection method, which shows that the model has stronger fine-grained classification ability under accurate positioning.

[0109] In addition, the global instance contrast effect of the detection model FCOS is also tested by the method, and the effect diagram can be seen in Figure 3 shown; from Figure 3It can be seen that many false detection results in the original model are corrected after the global instance comparison module is added, which shows that extracting class-specific features is very important to the classification accuracy of fine-grained detection. The above results also verify the effectiveness and universality of the method.

[0110] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A high-precision fine-grained SAR target detection method, characterized in that, The method comprises the following steps: S1, obtaining a SAR image to be detected and inputting the SAR image to be detected into a trained SAR image fine-grained detection model to output coordinates of all targets of interest in the SAR image to be detected and categories to which the targets of interest belong; S2, constructing a SAR image fine-grained detection model framework and inputting the training data set into the SAR image fine-grained detection model framework; extracting coordinates of targets of interest in the training data set and identifying categories of the extracted targets of interest by the SAR image fine-grained detection model framework; and combining the extracted and identified contents into a model collection data set; S3, randomly extracting instances from the global instance set to form a sampling data set; S4, respectively encoding targets of interest in the model collection data set and the sampling data set and performing global instance comparison; S5, optimizing the SAR image fine-grained detection model framework according to the comparison result until a final SAR image fine-grained detection model is obtained. The instance-level comparison loss used in the global instance comparison is defined as follows: The S1 specifically comprises the following steps: SAR image training samples are obtained; the SAR image training samples include a plurality of target regions; Wherein, L CTLi represents the contrast loss of each feature map; the numerator represents the sum of intra-class feature similarities; the denominator represents the sum of similarities of all features; f i and f j represent two different feature vectors; c i and c j represent the class labels of the corresponding vectors respectively; i and j represent the traversal indicating subscripts; cos < f i ,f j > represents the cosine similarity between two features; τ represents the temperature parameter; M represents the number of target regions in the first feature region set; N represents the number of target regions in the second feature region set; c i represents the known class label corresponding to each feature; L CTL represents the overall contrast loss obtained after averaging.

2. The high-precision fine-grained SAR target detection method of claim 1, wherein, The SAR image training samples and coordinates of all targets of interest in all target regions and categories to which the targets of interest belong are taken as a training data set; All target regions in the training data set are cropped and scaled to form a global instance set. In the S2, the SAR image fine-grained detection model framework extracts coordinates of targets of interest in the training data set and identifies categories of the extracted targets of interest, which specifically comprises the following steps: The backbone convolutional network in the SAR image fine-grained detection model framework extracts feature maps from the input SAR image training samples, and a plurality of feature maps of different sizes are obtained by calculating a feature pyramid; 3. The high-precision fine-grained SAR target detection method of claim 1, wherein, Each feature map is sent to a detection head to extract coordinates of targets of interest in each feature map and identify categories of the extracted targets of interest. For each feature map, the preparation work before global instance comparison comprises the following steps: According to the categories of the input targets of interest, the feature map is cropped into a plurality of regions, denoted as a first feature region set; 4. The high-precision fine-grained SAR target detection method of claim 3, wherein, Random sampling is performed on target regions in the global instance set, and the backbone convolutional network is used to extract features of the sampled target regions, denoted as a second feature region set; Global instance comparison is performed on targets of interest in the first feature region set and the second feature region set respectively after encoding. An encoder composed of average pooling and a fully connected layer is used to encode the targets of interest using tanh as a nonlinear activation function. The detection head uses a full convolution anchor-free structure; 5. The high-precision fine-grained SAR target detection method of claim 4, wherein, The detection head is used to predict the position and category confidence of each target of interest on feature maps of different scales.

6. The high-precision fine-grained SAR target detection method of claim 3, wherein, ​ ​ 7. The high-precision fine-grained SAR target detection method of claim 6, wherein, The detection head adopts a focus loss based on positioning quality perception as a classification loss; The focus loss based on positioning quality perception is expressed as: Wherein, q represents the intersection over union of the predicted bounding box and the real bounding box at the point; p represents the classification confidence of the current sample obtained by the SAR image fine-grained detection model; y=1 / 0 represents whether the point belongs to or does not belong to the category; when y=0, the corresponding q will also be set to 0; t, β, γ are all hyperparameters for controlling the proportion of each loss.

8. The high-precision fine-grained SAR target detection method of claim 1, wherein, The SAR image fine-grained detection model adopts a two-stage positioning method; In the first stage, the SAR image fine-grained detection model predicts the vector of all positive sample points, which respectively represents the rough distance prediction of the sample point to the four boundaries of the real bounding box; In the second stage, based on the prediction result of the first stage, a boundary perception bounding box fine optimization module is used to adjust the bounding box.

9. The high-precision fine-grained SAR target detection method of claim 1, wherein, The overall loss function of the SAR image fine-grained detection model is composed of three parts: classification loss, positioning loss and target positioning loss function.

Images