A lightweight SAR image ship target oblique frame detection method and system

Abstract

The present invention relates to a method and system for detecting oblique frame of ship targets in lightweight SAR images. The method comprises: constructing a lightweight backbone network; constructing a lightweight feature fusion structure at the output end of the lightweight backbone network; constructing a lightweight multi-task prediction branch structure at the output end of the lightweight feature fusion structure to obtain a first lightweight oblique frame detection model; training the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model; and inputting a SAR image of a ship target into the second lightweight oblique frame detection model to obtain the category and location of the ship target. The present invention improves detection speed while maintaining detection accuracy.

Description

Technical Field

[0001] The present invention relates to the fields of target detection and radar remote sensing, and in particular to a method and system for detecting a slant frame of a ship target in a lightweight SAR image. Background Art

[0002] In response to the problems of the current deep learning-based slant frame detection method for ship targets in SAR images, such as the large number of parameters, low efficiency, and the need for a pre-trained backbone network, the present invention discloses a lightweight SAR image ship slant frame detection method. The method is characterized by adding a slant frame parameter prediction branch based on the CenterNet framework. First, a lightweight backbone network is designed based on a de novo training strategy to extract features. Then, a lightweight feature fusion module is used to fuse deep semantic information and shallow spatial information. Finally, a lightweight multi-task prediction branch is used to predict the target center point, center point offset, and slant frame parameters. The entire model has few parameters and does not require a pre-trained backbone network, and can be trained directly from scratch. Experiments were conducted on the public SAR image ship target slant frame detection dataset SSDD+. The experimental results show that the proposed method improves the detection speed while maintaining detection accuracy, fully verifying the effectiveness of the proposed method. Summary of the Invention

[0003] The purpose of the present invention is to provide a lightweight SAR image ship target oblique frame detection method and system, which improves the detection speed while maintaining the detection accuracy.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A lightweight SAR image ship target oblique frame detection method, comprising:

[0006] Build a lightweight backbone network;

[0007] Constructing a lightweight feature fusion structure at the output end of the lightweight backbone network;

[0008] Constructing a lightweight multi-task prediction branch structure at the output end of the lightweight feature fusion structure to obtain a first lightweight oblique frame detection model;

[0009] Training the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model;

[0010] The SAR image of the ship target is input into the second lightweight oblique frame detection model to obtain the category and position of the ship target.

[0011] Optionally, the lightweight feature fusion structure is used to upsample the feature map output by the lightweight backbone network and fuse it with the shallow features to obtain deep semantic information and shallow detail information.

[0012] Optionally, the lightweight multi-task prediction branch structure includes a center point prediction branch, a center point offset prediction branch and a slant frame parameter prediction branch, the center point prediction branch is used to determine the target center point coordinates, the center point offset prediction branch is used to determine the target center point offset, and the slant frame parameter prediction branch is used to determine the slant frame parameters.

[0013] Optionally, the center point prediction branch adopts a Focal Loss loss function.

[0014] Optionally, the center point offset prediction branch and the slanted frame prediction branch adopt a Smooth L1 loss function.

[0015] Optionally, the lightweight backbone network consists of five stages: Stage 0, Stage 1, Stage 2, Stage 3 and Stage 4.

[0016] Optionally, the Stage 0 stage adopts a Stem structure, uses a 3×3 convolution with a stride of 2 instead of maximum pooling, and uses a 3×3 convolution instead of a large-size convolution.

[0017] Optionally, the Stage 1, Stage 2 and Stage 3 stages are composed of an OSA structure and maximum pooling. The OSA structure first aggregates the feature map through dense connections, then reduces the dimension through 1×1 convolution, and adds a channel attention module and residual connection. The maximum pooling downsamples the feature map by two times.

[0018] Optionally, the Stage 4 adopts the ASPP structure, first using dilated convolutions with different dilation coefficients to obtain multi-scale feature maps, then aggregating the multi-scale feature maps, and finally using 1×1 convolution to reduce the dimension of the feature maps.

[0019] A lightweight SAR image ship target oblique frame detection system, comprising:

[0020] Lightweight backbone network building module, used to build a lightweight backbone network;

[0021] A lightweight feature fusion structure construction module, used to construct a lightweight feature fusion structure at the output end of the lightweight backbone network;

[0022] A lightweight multi-task prediction branch structure construction module is used to construct a lightweight multi-task prediction branch structure at the output end of the lightweight feature fusion structure to obtain a first lightweight oblique frame detection model;

[0023] a training module, configured to train the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model;

[0024] The oblique frame detection module is used to input the SAR image of the ship target into the second lightweight oblique frame detection model to obtain the category and position of the ship target.

[0025] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0026] This paper proposes a lightweight method and system for detecting oblique frame of ship targets in SAR images. This method addresses the problems of deep learning-based methods for detecting oblique frame of ship targets in SAR images, which suffer from the high number of parameters, low efficiency, and the need for a pretrained backbone network. First, a lightweight backbone network is designed based on a de novo training strategy to extract features. A lightweight feature fusion module then fuses deep semantic information with shallow spatial information. Finally, a lightweight multi-task prediction branch predicts the target center point, center point offset, and oblique frame parameters. The entire model has few parameters and can be trained directly from scratch, eliminating the need for a pretrained backbone network. This method improves detection speed while maintaining detection accuracy, fully demonstrating its effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 This is a flow chart of the method for detecting oblique frames of ship targets in lightweight SAR images according to the present invention;

[0029] Figure 2 Training flow chart of an embodiment of the present invention;

[0030] Figure 3 A detection flow chart of an embodiment of the present invention;

[0031] Figure 4 This is a general structural diagram of a first lightweight oblique frame detection model according to an embodiment of the present invention;

[0032] Figure 5 1 is a block diagram of an OSP module according to an embodiment of the present invention;

[0033] Figure 6 It is a block diagram of the ASPP module according to an embodiment of the present invention;

[0034] Figure 7 is a schematic diagram of a slanted frame representation method according to an embodiment of the present invention;

[0035] Figure 8 2. Schematic diagram of offshore target detection results according to an embodiment of the present invention;

[0036] Figure 9 2. It is a schematic diagram of the nearshore target detection result according to an embodiment of the present invention;

[0037] Figure 10 This is a module diagram of the lightweight SAR image ship target oblique frame detection system of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] The purpose of the present invention is to provide a lightweight SAR image ship target oblique frame detection method and system, which improves the detection speed while maintaining the detection accuracy.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] A lightweight SAR image ship target oblique frame detection method, such as Figure 1 Shown, including:

[0042] Step 101: Build a lightweight backbone network.

[0043] The lightweight backbone network follows the principle of de novo training, utilizing the Stem architecture and densely connected OSA modules to obtain strong supervision information and accelerate network convergence. The final stage uses a multi-scale dilated convolution ASPP module to maintain resolution while acquiring a multi-scale receptive field. The network structure is shown in Table 1.

[0044] The input image resolution of the lightweight backbone network is 320×320. Based on a training strategy from scratch, Stem replaces max pooling with 3×3 convolutions with a stride of 2, replacing large-scale convolutions with 3×3 convolutions to reduce information loss. Stages 1-3 use the efficient densely connected OSA module to obtain strong supervision. The OSA module aggregates three layers of feature maps through dense connections, then reduces the dimensionality to the input dimension through 1×1 convolutions. A channel attention module and residual connections are then added. To preserve more spatial detail without compromising feature map resolution, Stage 4 uses the ASPP module. The ASPP module first uses dilated convolutions with different dilation coefficients to obtain multi-scale receptive fields without reducing resolution, then aggregates the multi-scale feature maps. Finally, 1×1 convolutions are used to reduce the feature maps to the input feature map dimension. Ultimately, the lightweight backbone network outputs a 128×20×20 feature map.

[0045] Step 102: Construct a lightweight feature fusion structure at the output end of the lightweight backbone network.

[0046] Among them, in order to restore the feature map output by the lightweight backbone network to high resolution, the lightweight feature fusion module first upsamples the feature map output by the backbone network and fuses it with the shallow features to obtain deep semantic information and shallow spatial information.

[0047] Specifically, the lightweight feature fusion module first upsamples the feature map output by the lightweight backbone network and fuses it with the shallow features to obtain deep semantic information and shallow detail information, and finally outputs a feature map with a size of 32×80×80.

[0048] Step 103: constructing a lightweight multi-task prediction branch structure at the output end of the lightweight feature fusion structure to obtain a first lightweight oblique frame detection model.

[0049] Among them, the lightweight multi-task prediction branch includes a center point prediction branch, a center point offset prediction branch, and a slant box parameter prediction branch, which predict the target center point coordinates, center point offset, and slant box parameters respectively.

[0050] Specifically, the center point prediction branch uses the coordinates of the target center point that are 4 times downsampled and rounded down as positive samples. Due to the small number of positive samples, in order to balance the positive and negative samples, the center point of the target is used as the center of the circle and the Gaussian kernel function is used to process the remaining positions of the center point prediction branch. The center point offset prediction branch predicts the offset. The oblique box representation method is as follows Figure 7As shown in the figure, the distances from eight edges to the center point are sampled at intervals of 1 / 2 within a half-cycle to represent the slanted box. Due to periodic symmetry, during the test phase, the slanted box parameters are equal to the predicted slanted box parameters in the other half-cycle. A total of 16 points represent the slanted box, and the minimum bounding box of these 16 points is obtained using OpenCV as the prediction result.

[0051] Step 104: training the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model.

[0052] The training process is as follows: Figure 2 As shown, there's no need to load model parameters pre-trained on a classification dataset. The entire network can be trained on the SSDD+ dataset after initialization using the Xavier function. During training, a loss function is used to calculate the error between the predicted and true values, and the model parameters are updated using gradient descent until the model converges. The final trained parameters are saved for the next stage of testing.

[0053] The loss function is a multi-task loss function. Since only the pixels at the target center point are positive samples in the center point prediction branch, and the remaining pixels are negative samples, to balance positive and negative samples, the center point prediction branch uses the Focal Loss loss function, while the center point offset prediction branch and the oblique box prediction branch use the Smooth L1 loss function.

[0054] Step 105: Input the SAR image of the ship target into the second lightweight oblique frame detection model to obtain the category and position of the ship target.

[0055] The detection process is as follows: Figure 3 As shown in the figure, first load the trained model parameters, then input the SAR image, and the detection result (category and position) of the target can be generated through calculation of the detection algorithm.

[0056] The overall structure of the first lightweight oblique frame detection model is as follows Figure 4 As shown in the figure, the algorithm consists of three main components: a lightweight backbone network, a lightweight feature fusion module, and a lightweight multi-task prediction branch module. The input image resolution is uniformly set to 320×320. First, the lightweight backbone network extracts features, outputting a 128×20×20 feature map. The lightweight feature fusion module then fuses deep semantic information with shallow spatial information, outputting a 32×80×80 feature map. Finally, a multi-task branch is added to the feature map to predict the center point coordinates H, the center point offset O, and the skew box parameter ρ.

[0057] Table 1 Lightweight backbone network

[0058]

[0059] The lightweight backbone network structure is shown in Table 1. The input image resolution is 320×320. Based on the training strategy from scratch, the Stem structure uses 3×3 convolution with a stride of 2 instead of maximum pooling, and uses 3×3 convolution instead of large-size convolution to reduce information loss; Stage 1-Stage 3 uses the efficient dense connection module OSA to obtain strong supervision information. Figure 5 As shown in the figure, the OSA module aggregates 3 layers of feature maps through dense connections, then reduces the dimension to the input dimension through 1×1 convolution, and adds a channel attention module and residual connection. In order to not reduce the resolution of the feature map and retain more spatial detail information, Stage 4 uses the ASPP module. Figure 6 As shown in the figure, the ASPP module first uses dilated convolutions with different dilation coefficients to obtain multi-scale receptive fields without reducing resolution. It then aggregates multi-scale feature maps and finally uses 1×1 convolution to reduce the feature maps to the input feature map dimension. Ultimately, the lightweight backbone network outputs a feature map of size 128×20×20.

[0060] like Figure 4 As shown in the figure, in order to restore the feature map output by the backbone network to high resolution, the lightweight feature fusion module first upsamples the feature map output by the backbone network and fuses it with the shallow features to obtain deep semantic information and shallow detail information, and finally outputs a feature map with a size of 32×80×80.

[0061] The lightweight multi-task prediction branch includes the center point prediction branch, the center point offset prediction branch and the slant box parameter prediction branch, which correspond to the center point, center point offset and slant box parameters respectively.

[0062] The center point prediction branch uses the coordinates of the target center point that are 4 times downsampled and rounded down. As a positive sample, where (c x , c y ) is the 4-fold downsampled coordinate of the target center point, Indicates rounding down. Since the number of positive samples is small, in order to balance the positive and negative samples, the center point of the target is used as the center of the circle, and the Gaussian kernel function is used to process the remaining positions (x, y) of the center point prediction branch, that is:

[0063]

[0064] where σ p is the adaptive scale factor, which is taken as 0.1 times of the short side of the target in this paper.

[0065] Since the coordinates of the center point of the oblique frame are rounded after being downsampled four times, there is a certain deviation from the actual center point coordinates. Used to represent the center point offset, and the branch prediction offset is predicted by the center point offset.

[0066] The oblique frame representation method is as follows Figure 7 As shown, within half a period [0, π) The distances from eight edges to the center point are sampled as intervals to represent the slant box. The final slant box is represented as ρ = (ρ1, ρ2, ..., ρ8). Due to periodic symmetry, during the test phase, the slant box parameter ρ′ is equal to ρ in the other half of the period. A total of 16 points represent the slant box. The minimum bounding box of these 16 points is obtained using OpenCV as the predicted slant box parameter.

[0067] like Figure 2 As shown in the figure, the lightweight multi-task prediction branch first uses 3×3 convolution to output a feature map of size 128×80×80, and then uses 1×1 convolution to reduce the dimension of the feature map to 1×80×80, 2×80×80, and 8×80×80, respectively predicting the center point coordinate H, center point offset O, and skew box parameter ρ.

[0068] Since only the pixels at the target center point in the center point prediction branch are positive samples and the rest of the pixels are negative samples, in order to balance the positive and negative samples, the center point prediction branch uses the FocalLoss loss function:

[0069]

[0070] Among them, α, β are penalty coefficients, and when the pixel point is a positive sample, p i =1, otherwise 0, Indicates the probability of predicting a positive sample.

[0071] The center point offset prediction branch uses Smooth L1 loss:

[0072]

[0073] Where o represents the actual offset of the target center point, Predict the offset for the target center point.

[0074] The oblique box prediction branch uses Smooth L1 loss:

[0075]

[0076] Where ρ is the true parameter of the slant box, is the prediction parameter.

[0077] The total loss function is as follows:

[0078] L=λ1L h +λ2L o +λ3L ρ (5)

[0079] Where N is the number of targets, λ1, λ2, λ3 are weight coefficients used to balance the proportion of different losses, and the oblique box parameter loss L ρ and center point offset loss L o Only positive samples are calculated, and the center point loss L h Calculate all samples.

[0080] During both training and testing, the input image resolution was uniformly set to 320×320, and the hyperparameters were set to α=2, β=4, λ1=1, λ2=1, and λ3=1. The Adam algorithm was used for optimization, with an initial learning rate of 0.001, which was then reduced using an exponential decay rule. The batch size was set to 32, and the number of iterations was set to 150. The experimental platform was a computer running a 64-bit Ubuntu 20.04 operating system, with an NVIDIA GTX1080Ti graphics card, accelerated using Cuda 11.0 and cuDNN 8.05. The programming language was Python, and the programming framework was Pytorch 1.7.1.

[0081] In the experiment, the model parameters, FLOPs, training speed, test speed and AP0.5 are used as evaluation indicators. FLOPs is the number of floating-point operations, and AP0.5 is the average accuracy when the IoU threshold is 0.5. AP is defined as follows:

[0082]

[0083] Among them, N td To detect the correct number of borders, N d is the number of all detected borders, N r is the number of true bounding boxes, P is the precision, R is the recall, and AP is the area under the PR curve.

[0084] To verify the effectiveness of our algorithm, we set up four different experiments to explore the impact of different structures on the detection results. The experimental results are shown in Table 2.

[0085] Table 2 Effects of different structures on test results

[0086]

[0087] Note: “-” indicates that the corresponding method is not adopted, and “√” indicates that the method is adopted.

[0088] As can be seen from the table, dense connection, ASPP, and feature fusion all have a certain impact on the detection results. When the number of parameters, FLOPs, and test time are similar, the highest AP is achieved by using all three structures at the same time.

[0089] To further validate the detection performance of our method, we compared it with typical recent slanted frame detection algorithms, including the PolarVector model and the BBAVector model. We used the best settings of these models from the literature and compared them with the best settings of our method. The results are shown in Table 3, which summarizes the number of parameters, FLOPs, test time, and AP of the different models.

[0090] Table 3 Comparison of detection results of different models

[0091] method Parameter quantity FLOPS Testing Time AP Pre-training PolarVector 71.73M 189.99G 78ms 89.46％ √ BBAVector 72.42M 203.70G 59ms 88.88％ √ Methods 1.32M 2.00G 33ms 79.82％ -

[0092] As can be seen from the table, compared with other methods, the proposed method greatly reduces the number of parameters, FLOPs and test time while maintaining a certain AP.

[0093] The algorithm of this invention selects SAR images of different resolutions in nearshore and offshore scenes to display the detection results. Figure 8 This is the offshore scene detection result. Figure 9 The detection results are for nearshore scenes. The detection results show that the model is applicable to multi-resolution images and multi-scale ship target detection in multiple scenes.

[0094] Based on the above method, the present invention also provides a lightweight SAR image ship target oblique frame detection system, such as Figure 10 Shown, including:

[0095] A lightweight backbone network construction module 201 is used to construct a lightweight backbone network;

[0096] A lightweight feature fusion structure construction module 202 is used to construct a lightweight feature fusion structure at the output end of the lightweight backbone network;

[0097] A lightweight multi-task prediction branch structure construction module 203 is used to construct a lightweight multi-task prediction branch structure at the output end of the lightweight feature fusion structure to obtain a first lightweight oblique frame detection model;

[0098] A training module 204 is configured to train the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model;

[0099] The oblique frame detection module 205 is used to input the SAR image of the ship target into the second lightweight oblique frame detection model to obtain the category and position of the ship target.

[0100] The present invention also discloses the following technical effects:

[0101] The present invention discloses a lightweight SAR image ship target oblique frame detection method and system. This method addresses the problems of deep learning-based SAR image ship target oblique frame detection methods, such as large number of parameters, low efficiency, and the need for pre-trained backbone networks. First, a lightweight backbone network is designed based on a de novo training strategy to extract features. Then, a lightweight feature fusion module is used to fuse deep semantic information and shallow spatial information. Finally, a lightweight multi-task prediction branch is used to predict the target center point, center point offset, and oblique frame parameters. The entire model has few parameters and does not require a pre-trained backbone network; it can be trained directly from scratch. Experiments were conducted on a public SAR image ship target detection dataset. The experimental results show that the proposed method improves detection speed while maintaining detection accuracy, fully verifying the effectiveness of the proposed method.

[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0103] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A lightweight SAR image ship target oblique frame detection method, characterized by: include: Build a lightweight backbone network. Based on the training strategy from scratch, Stem uses convolution instead of maximum pooling. Stage1-Stage3 use OSA module to aggregate feature maps through dense connections; Stage 4 uses the ASPP module to obtain multi-scale receptive fields using dilated convolutions with different dilation coefficients without reducing the resolution. Finally, the lightweight backbone network outputs a feature map of size 128×20×20. Constructing a lightweight feature fusion structure at the output end of the lightweight backbone network; A lightweight multi-task prediction branch structure is constructed at the output end of the lightweight feature fusion structure to obtain a first lightweight slant frame detection model; the lightweight multi-task prediction branch structure includes a center point prediction branch, a center point offset prediction branch, and a slant frame parameter prediction branch, the center point prediction branch is used to determine the target center point coordinates, the center point offset prediction branch is used to determine the target center point offset, and the slant frame parameter prediction branch is used to determine the slant frame parameters; the center point prediction branch adopts a Focal Loss loss function; the center point offset prediction branch and the slant frame parameter prediction branch adopt a Smooth L1 loss function; Training the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model; The SAR image of the ship target is input into the second lightweight oblique frame detection model to obtain the category and position of the ship target.

2. The method for detecting oblique frames of ship targets in lightweight SAR images according to claim 1, characterized in that: The lightweight backbone network consists of five stages: Stage 0, Stage 1, Stage 2, Stage 3 and Stage 4.

3. The method for detecting oblique frames of ship targets in lightweight SAR images according to claim 1, characterized in that: The lightweight feature fusion structure is used to upsample the feature map output by the lightweight backbone network and fuse it with the shallow features to obtain deep semantic information and shallow detail information.

4. The method for detecting oblique frames of ship targets in lightweight SAR images according to claim 2, characterized in that: The Stage 0 stage adopts the Stem structure, uses a 3×3 convolution with a step size of 2 instead of maximum pooling, and uses a 3×3 convolution instead of a large-size convolution.

5. The method for detecting oblique frames of ship targets in lightweight SAR images according to claim 2, characterized in that: The Stage 1-Stage 3 stages are composed of the OSA structure and maximum pooling. The OSA structure first aggregates the feature map through dense connections, then reduces the dimension through 1×1 convolution, and adds a channel attention module and residual connection. The maximum pooling downsamples the feature map by two times.

6. The method for detecting oblique frames of ship targets in lightweight SAR images according to claim 2, characterized in that: The Stage 4 adopts the ASPP structure, first using dilated convolution with different dilation coefficients to obtain multi-scale feature maps, then aggregate the multi-scale feature maps, and finally use 1×1 convolution to reduce the dimension of the feature maps.

7. A lightweight SAR image ship target oblique frame detection system, characterized by: include: A lightweight backbone network building module is used to build a lightweight backbone network. Based on the training strategy from scratch, Stem uses convolution instead of maximum pooling. Stages 1-3 use the OSA module to aggregate feature maps through dense connections. Stage 4 uses the ASPP module to obtain multi-scale receptive fields using dilated convolutions with different dilation coefficients without reducing resolution. Finally, the lightweight backbone network outputs a feature map of size 128×20×20. A lightweight feature fusion structure construction module, used to construct a lightweight feature fusion structure at the output end of the lightweight backbone network; A lightweight multi-task prediction branch structure construction module is used to construct a lightweight multi-task prediction branch structure at the output end of the lightweight feature fusion structure to obtain a first lightweight slant frame detection model; the lightweight multi-task prediction branch structure includes a center point prediction branch, a center point offset prediction branch and a slant frame parameter prediction branch, the center point prediction branch is used to determine the target center point coordinates, the center point offset prediction branch is used to determine the target center point offset, and the slant frame parameter prediction branch is used to determine the slant frame parameters; the center point prediction branch adopts a FocalLoss loss function; the center point offset prediction branch and the slant frame parameter prediction branch adopt a Smooth L1 loss function; A training module, configured to train the first lightweight oblique frame detection model to obtain a second lightweight oblique frame detection model; The oblique frame detection module is used to input the SAR image of the ship target into the second lightweight oblique frame detection model to obtain the category and position of the ship target.

Images