A ship target instance segmentation method based on Gaussian coding and decoding

Abstract

The application discloses a ship target instance segmentation method based on Gaussian coding and decoding, combines a center point extraction network based on a Gaussian heat map with a feature fusion method based on Gaussian feature coding, designs an anchor frame-free instance segmentation method for encoding ship instances into two-dimensional elliptical Gaussian distribution, and realizes high-precision synthetic aperture radar (SAR) image ship target instance segmentation. The former can effectively avoid the problem that adjacent ship targets are difficult to distinguish due to the vertical frame caused by encoding SAR image ship instances into the vertical frame by using a Mask R-CNN method on the basis of improving the ship instance coding efficiency. The latter introduces Gaussian features into the mask segmentation process, and effectively improves the discrimination ability of a mask head for different ship instances and ship boundaries.

Description

Technical Field

[0001] This invention belongs to the field of SAR image processing and relates to a method for segmenting ship target instances based on Gaussian encoding and decoding. Background Technology

[0002] Synthetic Aperture Radar (SAR) imagery for ship target detection is an important branch of the SAR field. Leveraging its all-weather, all-time imaging advantages, it plays a crucial role in maritime traffic surveillance, fisheries monitoring, maritime vessel management, and intelligence gathering. With the development of deep learning technology, SAR ship target detection algorithms based on Convolutional Neural Networks (CNNs) have significantly improved in terms of detection accuracy, speed, and robustness. However, they still have certain shortcomings in accuracy and mask segmentation fineness. Therefore, instance segmentation algorithms that can better locate ship positions and segment ship masks have received increasing attention.

[0003] Widely used methods in instance segmentation, such as Mask R-CNN, Yolact, and CondInst, encode instance location information into vertical bounding boxes and employ non-maximum suppression (NMS) for redundancy removal. Synthetic aperture radar (SAR) images of ships exhibit significant variations in aspect ratio and orientation, and often display densely packed targets. This leads to difficulties in encoding the shape of ship instances and the possibility of misidentifying multiple instances as a single instance during redundancy removal. To address these issues, encoding methods that better fit the ship's shape and more efficient ship coordinate decoding methods are needed.

[0004] Meanwhile, instance segmentation methods are limited by the receptive field of the convolution kernel during mask segmentation, making it difficult to distinguish the positions of different instances and differentiate between foreground and background, easily leading to oversegmentation or undersegmentation. To address this issue, it is necessary to introduce feature information richer in ship shape and position characteristics into the mask feature map. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a ship target instance segmentation method based on Gaussian encoding and decoding. This method uses a multi-scale deep convolutional neural network to extract high-resolution feature maps, generates a ship Gaussian heatmap through a Gaussian encoding and decoding mask segmenter, obtains the ship's center position through Gaussian center point decoding, and combines a feature fusion method based on Gaussian feature encoding. It uses an instance segmentation technique based on dynamic convolution to perform mask segmentation on SAR ship instances, which has high segmentation accuracy and good generalizability.

[0006] The technical solution adopted by this invention to solve its technical problem is: a method for segmenting ship target instances based on Gaussian encoding and decoding, comprising the following steps:

[0007] (1) Set a certain number of iterations, and process the images in the training set sequentially during each iteration;

[0008] (2) Input images from the training set and extract high-resolution feature maps using a multi-scale deep convolutional neural network;

[0009] (3) Input the obtained feature map F into the Gaussian encoder-decoder mask segmenter to encode the ship instance into a two-dimensional elliptical Gaussian distribution, and combine the Gaussian center point decoding method with the feature fusion method based on Gaussian feature encoding to perform ship mask segmentation.

[0010] (4) Supervised training was performed on the Gaussian heatmap and ship mask segmentation results, and a loss function was set to guide the network regression.

[0011] (5) Return to step (2) and continue to input the next image from the training set until all images have been processed and the current iteration process ends;

[0012] (6) Continue the next iteration process until the set number of iterations is met, the training process ends, and the network parameters are determined;

[0013] (7) Perform instance segmentation on the images in the test set using the trained neural network and output the segmentation results;

[0014] Furthermore, the specific process of extracting high-resolution feature maps from the input images in the training set using a multi-scale deep convolutional neural network in step (2) is as follows:

[0015] The input SAR image is first reduced in resolution by passing it through two 3×3 convolutional layers with a stride of 1 and 2x downsampling. Then, a feature extraction network extracts features at four different scales, from concrete to abstract, with the feature map size reduced to half its original size at each scale. During feature extraction at different scales, a hierarchical deep aggregation method is used to aggregate feature maps at the same scale into a single high-semantic feature map. After obtaining the feature maps at four scales, an iterative deep aggregation method is used to aggregate the feature maps at different scales into the highest-resolution feature map, ultimately obtaining a high-resolution, high-semantic-information feature map F.

[0016] Furthermore, in step (3), the obtained feature map F is input into the Gaussian encoder-decoder mask segmenter to encode the ship instance as a two-dimensional elliptical Gaussian distribution. The ship mask segmentation is performed by combining the Gaussian center point decoding method with the feature fusion method based on Gaussian feature encoding. The specific process is as follows:

[0017] Furthermore, the Gaussian encoder-decoder mask segmenter comprises three parts: a Gaussian feature encoder, a Gaussian center point encoder-decoder, and an instance segmentation network. First, the feature map F is input into the Gaussian feature encoder. The Gaussian feature encoder obtains spatial and channel attention from the feature map through spatial average pooling and channel average pooling. Then, it passes through a fully convolutional encoder-decoder composed of 1×1 convolutional kernels to obtain two sets of attention weights. The feature map is then weighted through element-wise multiplication to obtain two sets of feature maps F. gauss and F mask The two sets of feature maps obtained by this attention-based feature map encoder can achieve decoupling of spatially relevant locations, and are respectively applicable to Gaussian heatmap encoding tasks that focus on the center location of ship instances and instance segmentation tasks that focus on the edge locations of ship instances.

[0018] Next, the feature map F gauss Input Gaussian center point codec, first for F gauss A 2x upsampling process is performed, followed by a Gaussian heatmap obtained through a 3×3 convolutional layer and sigmoid encoding. Then, center point decoding is performed on the Gaussian heatmap: first, a threshold T is set, and Gaussian heat values ​​less than T are set to 0; then, local peak points are identified as candidate center points; finally, based on the following four criteria, it is determined whether the local peak points originate from the same ship instance target, and duplicate candidate center points are removed:

[0019] ① If there is a point with a heat value of 0 on the line segment between two points, then the two points belong to different targets;

[0020] ②If the difference between the heat value of the candidate center point and the minimum heat value on the line segment between the two points is less than M, then the two points belong to the same target;

[0021] ③ If the average heat value of the line segment between two points is greater than the heat value of the candidate center point, then the two points belong to the same target;

[0022] ④ If the above three conditions are not met, then the two points belong to different objectives;

[0023] The center point after deduplication is proposed as the center point of the ship example. proposal , feature map F mask F and Center Point Proposal proposal The instance segmentation network is used for ship mask segmentation. The instance segmentation network is based on a dynamic mask head instance segmentation network; the input feature map F yields the dynamic mask head parameter feature map F. param and mask feature map Then F mask and The features are fused by addition to obtain a mask feature map rich in the boundary features of ship instances. Finally, dynamic mask head parameters are extracted based on the center point proposal in the mask feature map. The ship's mask is segmented on the surface.

[0024] Furthermore, in step (4), the Gaussian heatmap and ship mask segmentation results are subjected to supervised training, and a loss function is set to guide network regression. The specific process is as follows:

[0025] Based on the true value of the ship mask, the ship mask is encoded as a two-dimensional elliptic Gaussian distribution. First, the mean and covariance matrix of the two-dimensional Gaussian distribution are calculated.

[0026]

[0027]

[0028] Where (x, y) are the coordinates of different positions, Mask ship For ship binary masking, g x,y Let be the gray value of the image at coordinates (x, y), and u and ∑ be the calculated Gaussian mean and covariance. Then, the two-dimensional Gaussian distribution is calculated using the empirical formula for the two-dimensional Gaussian distribution.

[0029]

[0030] Here, k is a hyperparameter that adjusts the range of the Gaussian heatmap distribution. To ensure that the boundary positions of adjacent ship instances are given more attention during the supervised Gaussian heatmap regression process, boundary weighting is required.

[0031]

[0032]

[0033]

[0034] Where, d i (x,y) represents the nearest distance between coordinates (x,y) and the mask of the i-th ship, and D(x,y) is the set of nearest distances between coordinates (x,y) and all ship masks in the SAR image. is the binary mask of the i-th ship mask, N is the total number of ships in the SAR image, and k is a hyperparameter for adjusting the range of Gaussian heat distribution.

[0035] The difference between the Gaussian heatmap and the ground truth is determined by the difference between the Gaussian heatmap and the ground truth. border The weighted mean square error is used to represent this:

[0036]

[0037] Where H and W are the length and width of the feature map, respectively. The mask segmentation results are supervised using the most common DiceLoss algorithm, specifically represented as follows:

[0038]

[0039] Among them, Mask gt and Mask pred These represent the true mask value and the predicted mask result, respectively.

[0040] The loss function of the entire network is composed of the above two parts through a weighted sum, and is expressed as:

[0041] Loss total =α1Loss gauss +α2Loss mask

[0042] Here, α1 and α2 are hyperparameters of the loss function. Setting α1 = 100 and α2 = 1 makes the loss values ​​of each part uniform to the same order of magnitude, ensuring similar convergence speeds.

[0043] Furthermore, in step (7), the trained neural network is used to perform instance segmentation on the images in the test set and output the segmentation results. The specific process is as follows:

[0044] First, an image is input from the test set, and a high-resolution feature map F is extracted using a pre-trained multi-scale feature extraction network. Next, the high-resolution feature map is input into a Gaussian encoder-decoder mask segmenter, which generates two sets of feature maps F suitable for Gaussian regression and mask segmentation, respectively. gauss and F mask , will F gauss Input a Gaussian center point encoder / decoder to perform center point prediction, and obtain the center point proposal (Center). proposal Finally, the Center proposal F mask The high-resolution feature map F is input into the instance segmentation network to perform dynamic mask head ship instance segmentation based on feature fusion, and the segmentation result is output.

[0045] The principle of this invention is as follows: Based on Gaussian encoding and decoding, a model is constructed for SAR image ship target instance segmentation. First, based on an attention mechanism, a high-resolution feature map is encoded into two sets of Gaussian feature maps, which are used for Gaussian regression and mask segmentation, respectively. Then, the feature map suitable for Gaussian regression is encoded into a two-dimensional elliptical Gaussian distribution of ship instances, and the Gaussian center point is extracted according to the method proposed in this invention. Finally, the Gaussian-encoded mask feature map and the original mask feature map are fused, and dynamic mask segmentation is performed based on the coordinates of the Gaussian center point to complete high-precision instance segmentation of SAR images of ships.

[0046] The advantages of this invention compared to the prior art are:

[0047] It effectively avoids the problems of methods like Mask R-CNN encoding ship instances as vertical boxes, which is difficult to encode and detects adjacent instances as one instance during nonmaximum suppression. The feature map fusion of Gaussian encoding introduces more feature information rich in ship shape and position features into the mask segmentation process, effectively improving the mask head's ability to distinguish different ship instances and ship boundaries. Attached Figure Description

[0048] Figure 1 This is a flowchart of a ship target instance segmentation method based on Gaussian encoding and decoding; where, Figure 1 (a) is a flowchart of the training process. Figure 1 (b) is a flowchart of the testing process;

[0049] Figure 2 This is a diagram illustrating the multi-scale feature extraction process.

[0050] Figure 3 Diagram of a Gaussian encoder / decoder mask splitter;

[0051] Figure 4 This is a diagram of the Gaussian feature encoder structure.

[0052] Figure 5 Diagram of the Gaussian center point encoder / decoder structure;

[0053] Figure 6 Schematic diagram of instance segmentation network;

[0054] Figure 7 A boundary position weight map for adjacent ship instances;

[0055] Figure 8 This is a segmentation result image for a typical scenario. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, the specific implementation steps of the ship target instance segmentation method based on Gaussian encoding and decoding of the present invention are as follows:

[0058] (1) Set a certain number of iterations, and process the images in the training set sequentially during each iteration;

[0059] (2) Input images from the training set and extract high-resolution feature maps using a multi-scale deep convolutional neural network. The specific process is as follows:

[0060] like Figure 2As shown, the input image is a SAR image. First, the resolution of the input SAR image is reduced by two 3×3 convolutional layers with a stride of 1 and double downsampling. Then, a feature extraction network extracts features at four different scales from concrete to abstract, with the size of the feature map at each scale being reduced to half its original size. When extracting features at different scales, a hierarchical deep aggregation method is used to aggregate feature maps at the same scale into a single high-semantic feature map. After obtaining the feature maps at four scales, an iterative deep aggregation method is used to aggregate the feature maps at different scales into the feature map with the highest resolution, ultimately obtaining a high-resolution feature map F with high semantic information.

[0061] (3) Input the obtained feature map F into the Gaussian encoder-decoder mask segmenter to encode the ship instance into a two-dimensional elliptical Gaussian distribution. Combine the Gaussian center point decoding method with the feature fusion method based on Gaussian feature encoding to perform ship mask segmentation. The specific process is as follows:

[0062] like Figure 3 As shown, the Gaussian encoder-decoder mask segmenter consists of three parts: a Gaussian feature encoder, a Gaussian center point encoder-decoder, and an instance segmentation network. First, the feature map F is input into the Gaussian feature encoder, as shown below. Figure 4 As shown, the Gaussian feature encoder obtains spatial and channel attention from the feature map through spatial average pooling and channel average pooling. Then, it obtains two sets of attention weights through a fully convolutional encoder and decoder composed of 1×1 convolutional kernels. The feature map is then weighted by element-wise multiplication to obtain two sets of feature map F. gauss and F mask The two sets of feature maps obtained by this attention-based feature map encoder can achieve decoupling of spatially relevant locations, and are respectively applicable to Gaussian heatmap encoding tasks that focus on the center location of ship instances and instance segmentation tasks that focus on the edge locations of ship instances.

[0063] Next, the feature map F gauss Input Gaussian center point codec, such as Figure 5 As shown, firstly for F gauss A 2x upsampling process is performed, followed by a Gaussian heatmap obtained through a 3×3 convolutional layer and sigmoid encoding. Then, center point decoding is performed on the Gaussian heatmap: first, a threshold T is set, and Gaussian heat values ​​less than T are set to 0; then, local peak points are identified as candidate center points; finally, based on the following four criteria, it is determined whether the local peak points originate from the same ship instance target, and duplicate candidate center points are removed:

[0064] ① If there is a point with a heat value of 0 on the line segment between two points, then the two points belong to different targets;

[0065] ②If the difference between the heat value of the candidate center point and the minimum heat value on the line segment between the two points is less than M, then the two points belong to the same target;

[0066] ③ If the average heat value of the line segment between two points is greater than the heat value of the candidate center point, then the two points belong to the same target;

[0067] ④ If the above three conditions are not met, then the two points belong to different objectives;

[0068] The center point after deduplication is proposed as the center point of the ship example. proposal , feature map F mask F and Center Point Proposal proposal Input an instance segmentation network for ship mask segmentation. The instance segmentation network is based on a dynamic mask head instance segmentation network, such as... Figure 6 As shown, the input feature map F yields the dynamic mask head parameter feature map F. param and mask feature map Then F mask and The features are fused by addition to obtain a mask feature map rich in the boundary features of ship instances. Finally, dynamic mask head parameters are extracted based on the center point proposal in the mask feature map. The ship's mask is segmented on the surface;

[0069] (4) Supervised training is performed on the Gaussian heatmap and ship mask segmentation results. A loss function is set to guide the network regression. The specific process is as follows:

[0070] Based on the true value of the ship mask, the ship mask is encoded as a two-dimensional elliptic Gaussian distribution. First, the mean and covariance matrix of the two-dimensional Gaussian distribution are calculated.

[0071]

[0072]

[0073] Where (x, y) are the coordinates of different positions, Mask ship For ship binary masking, g x,y Let (x, y) be the gray value of the image at coordinates (x, y). Then, calculate the two-dimensional Gaussian distribution using the empirical formula for the two-dimensional Gaussian distribution.

[0074]

[0075] Where k is a hyperparameter that adjusts the range of the Gaussian heatmap distribution. To ensure that the boundary positions of adjacent ship instances are given more attention during the supervised Gaussian heatmap regression process, it is necessary to calculate the boundary weights and generate a result such as... Figure 7 The weighting graph shown.

[0076]

[0077]

[0078]

[0079] Where, d i (x,y) represents the nearest distance between coordinates (x,y) and the mask of the i-th ship, and D(x,y) is the set of nearest distances between coordinates (x,y) and all ship masks in the SAR image. is the binary mask of the i-th ship mask, N is the total number of ships in the SAR image, and k is a hyperparameter for adjusting the range of Gaussian heat distribution.

[0080] The difference between the Gaussian heatmap and the ground truth is determined by the difference between the Gaussian heatmap and the ground truth. border The weighted mean square error is used to represent this:

[0081]

[0082] Where H and W are the length and width of the feature map, respectively. The mask segmentation results are supervised using the most common Dice Loss algorithm, specifically represented as follows:

[0083]

[0084] Among them, Mask gt and Mask pred These represent the true mask value and the predicted mask result, respectively.

[0085] The loss function of the entire network is composed of the above two parts through a weighted sum, and is expressed as:

[0086] Loss total =α1Loss gauss +α2Loss mask

[0087] Here, α1 and α2 are hyperparameters of the loss function. Setting α1 = 100 and α2 = 1 makes the loss values ​​of each part uniform to the same order of magnitude, ensuring similar convergence speeds.

[0088] (5) Return to step (2) and continue to input the next image from the training set until all images have been processed and the current iteration process ends;

[0089] (6) Continue the next iteration process until the set number of iterations is met, the training process ends, and the network parameters are determined;

[0090] (7) Using the trained neural network, perform instance segmentation on the images in the test set and output the segmentation results. The specific process is as follows:

[0091] First, an image is input from the test set, and a high-resolution feature map F is extracted using a pre-trained multi-scale feature extraction network. Next, the high-resolution feature map is input into a Gaussian encoder-decoder mask segmenter, which generates two sets of feature maps F suitable for Gaussian regression and mask segmentation, respectively. gauss and F mask , will F gauss Input a Gaussian center point encoder / decoder to perform center point prediction, and obtain the center point proposal (Center). proposal Finally, the Center proposal F mask The high-resolution feature map F is input into the instance segmentation network to perform dynamic mask head ship instance segmentation based on feature fusion, and the segmentation result is output.

[0092] Figure 8 The segmentation results are shown in some typical scenarios. As can be seen from the segmentation results, regardless of whether it is on the sea surface or near the shore, and regardless of whether the ships are sparsely or densely arranged, this method can generate instance masks relatively smoothly with a small number of false alarms and missed detections, and suppress over-segmentation or under-segmentation of the mask.

[0093] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0094] Although preferred embodiments and accompanying drawings have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the technical solutions protected by this invention should not be limited to those disclosed in the preferred embodiments and accompanying drawings.

Claims

1. A method for segmenting ship target instances based on Gaussian encoding and decoding, characterized in that, The method includes the following steps: Step (1) Set a certain number of iterations, and process the images in the training set sequentially during each iteration; Step (2) Select images from the training set as input images and extract high-resolution feature maps using a multi-scale deep convolutional neural network. ; Step (3) Obtain the feature map Input a Gaussian encoder-decoder mask segmenter to encode ship instances into a two-dimensional elliptical Gaussian distribution. Combine the Gaussian center point decoding method with the feature fusion method based on Gaussian feature encoding to perform ship mask segmentation, and obtain the Gaussian heat map and ship mask segmentation results. Step (4) Supervised training is performed on the Gaussian heatmap and ship mask segmentation results, and a loss function is set to guide the network regression; Step (5) returns to step (2) and continues to select images from the training set as the input for the next image until all images have been processed, at which point the current iteration ends. Step (6) Continue the next iteration process until the set number of iterations is met, the training process ends, and the network parameters are determined; Step (7) Use the trained neural network to perform instance segmentation on the images in the test set and output the segmentation results; The specific implementation process of step (3) is as follows: the Gaussian encoder-decoder mask segmenter includes three parts: Gaussian feature encoder, Gaussian center point encoder-decoder and instance segmentation network; First, the feature map The input is a Gaussian feature encoder, which obtains spatial and channel attention for the feature map through spatial average pooling and channel average pooling. Then, it passes through a fully convolutional encoder / decoder composed of 1×1 convolutional kernels to obtain two sets of attention weights. Finally, element-wise multiplication is used to weight the feature map, resulting in two sets of feature maps. and ; Next, the feature map The input Gaussian center point encoder-decoder first processes the feature map. Perform a 2x upsampling, then obtain the Gaussian heatmap using a 3×3 convolutional layer and sigmoid encoding. Then, the Gaussian heatmap Perform center point decoding, which specifically involves: first, setting a threshold. It will be less than the threshold. The Gaussian heat value is set to 0. Then, local peak points are searched as candidate center points. Finally, based on the following four criteria, it is determined whether the local peak points originate from the same ship instance target, and duplicate candidate center points are deduplicated: ① If there is a point with a heat value of 0 on the line segment between two points, then the two points belong to different targets; ② The difference between the heat value of the candidate center point and the minimum heat value on the line segment between the two points is less than If so, then the two points belong to the same target; ③ If the average heat value of the line segment between two points is greater than the heat value of the candidate center point, then the two points belong to the same target; ④ If the above three conditions are not met, then the two points belong to different objectives; The proposed center point after deduplication is used as the center point of the ship example. , feature map , and central point proposal The input instance segmentation network performs ship mask segmentation. The instance segmentation network is based on a dynamic mask head instance segmentation network, and the input feature map is... Obtain the dynamic mask head parameter feature map and mask feature map Then and The features are fused by addition to obtain a mask feature map rich in the boundary features of ship instances. Finally, dynamic mask head parameters are extracted based on the center point proposal, and then applied to the mask feature map. The ship's mask is segmented on the surface.

2. The method according to claim 1, characterized in that, The specific implementation process of step (2) is as follows: The input image is a SAR image. First, the resolution of the input SAR image is reduced by two 3×3 convolutional layers with a stride of 1 accompanied by double downsampling. Then, a feature extraction network is used to extract four feature maps at different scales from specific to abstract. The size of each feature map is reduced to half of its original size. When extracting features at different scales, a hierarchical deep aggregation method is used to aggregate feature maps at the same scale into a high semantic feature map. After obtaining the feature maps at four scales, an iterative deep aggregation method is used to aggregate the feature maps at different scales into the feature map with the highest resolution. Finally, a high-resolution feature map with high semantic information is obtained. .

3. The method according to claim 2, characterized in that, The specific implementation process of step (4) is as follows: based on the true value of the ship mask, the ship mask is encoded as a two-dimensional elliptic Gaussian distribution. First, the mean and covariance matrix of the two-dimensional Gaussian distribution are calculated. in, For the coordinates of different positions, For ship binary masking, coordinates The grayscale value of the image. and To calculate the Gaussian mean and covariance, the two-dimensional Gaussian distribution is then calculated using the empirical formula for the two-dimensional Gaussian distribution. Where k is a hyperparameter that adjusts the range of the Gaussian heat distribution, and the boundary weighting is calculated. in, Representing coordinates The closest distance to the shield of the i-th ship. Coordinates The set of closest distances to all ship masks in SAR images. is the binary mask of the i-th ship's mask, and N is the total number of ships in the SAR image. Gaussian heatmap The difference from the true value is due to the passage The weighted mean square error is used to represent this: Where H and W are the length and width of the feature map, respectively. The mask segmentation result is supervised using the most common Dice Loss, specifically as follows: in, and These represent the true mask value and the predicted mask result, respectively. The loss function of the entire network is composed of the above two parts through a weighted sum, and is expressed as: in, and Set the hyperparameters of the loss function. , This ensures that the loss values ​​of each part are unified to the same order of magnitude, guaranteeing similar convergence speeds.

4. The method according to claim 1, characterized in that, The specific implementation process of step (7) is as follows: First, input images from the test set, and use the trained multi-scale feature extraction network to extract high-resolution feature maps of the images. Next, the high-resolution feature map is input into the Gaussian encoder-decoder mask segmenter, which then generates two sets of feature maps suitable for Gaussian regression and mask segmentation, respectively. and ,Will Input a Gaussian center point encoder / decoder to perform center point prediction, and obtain center point proposals. Finally , and high-resolution feature maps Input an instance segmentation network, perform dynamic mask head ship instance segmentation based on feature fusion, and output the segmentation results.

Images