A sar image ship instance segmentation method based on global semantic boundary attention network

Abstract

The application discloses a SAR ship instance segmentation method based on a global semantic boundary attention network, and aims at solving the problem of limited target frame positioning capability in the prior art. The application is based on the deep learning theory and mainly comprises a global context information modeling module and a boundary attention prediction module. The global context information modeling module establishes a long-distance dependency relationship by enhancing the semantic information of features for multiple times, thereby effectively reducing background interference. The boundary attention prediction module improves the positioning capability of the target frame by predicting the boundary information of the target twice. The method provided by the application is superior to other SAR ship instance segmentation methods based on the deep learning in average precision (AP). The application can overcome the problem of limited target frame positioning capability in the prior art, and improve the instance segmentation precision of ships in SAR images.

Description

Technical Field

[0001] This invention belongs to the field of Synthetic Aperture Radar (SAR) image interpretation technology, and relates to a SAR image ship instance segmentation method based on a global semantic boundary attention network. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an excellent sensor. It provides high-resolution images by measuring the radar scattering characteristics of targets, unaffected by light or weather conditions, and is widely used in surveying, transportation, oceanography, remote sensing, and other fields. Ship surveillance, beneficial for disaster relief, traffic control, and fisheries monitoring, is currently a hot research topic. Compared to optical, infrared, and hyperspectral sensors, SAR is more adaptable to marine climate change environments and is better suited for ship surveillance. Therefore, the use of SAR for ship surveillance is receiving increasing attention.

[0003] Traditional methods typically rely on expert experience to manually create features, which is time-consuming and labor-intensive, limiting wider application. In recent years, deep learning-based detection, classification, and recognition methods have rapidly developed across various fields, particularly in pedestrian detection, face recognition, image classification, and speech translation, where they have seen in-depth and widespread applications. Introducing deep learning into SAR vessel instance segmentation has significant application potential. More and more scholars are conducting research on deep learning-based SAR vessel instance segmentation methods. For example, Su Hao et al. used a convolutional neural network-based model for remote sensing image instance segmentation, but they did not consider the features of SAR vessels, affecting further accuracy improvements. Gao Fei et al. proposed an anchor-free instance segmentation network, but this model cannot handle complex scenes and cases. Zhao Danpei et al. proposed a collaborative attention-based SAR vessel instance segmentation method, but their method still missed many small vessels and near-shore vessels. In general, most existing SAR vessel instance segmentation methods have limited target box localization capabilities, leaving room for further improvement in segmentation accuracy.

[0004] Therefore, to address this issue, this paper proposes a SAR image ship instance segmentation method based on a global semantic boundary attention network, which improves instance segmentation accuracy by enhancing the localization capability of the target bounding boxes. This method mainly comprises two modules to improve the localization capability of the target bounding boxes. The first module is a global context information modeling module, consisting of a content-aware feature reorganization subnetwork, a multi-view domain feature extraction subnetwork, and a global feature self-attention subnetwork connected in series. The global context information modeling module models the long-distance dependencies of the ship's surrounding environment through a larger view domain, effectively reducing background interference and extracting more discriminative regional features. The second module is a boundary attention prediction module, consisting of a boundary attention feature extraction subnetwork, a boundary coarse localization subnetwork, a boundary fine localization subnetwork, and a boundary-guided classification re-scoring subnetwork connected in series. Unlike traditional boundary regression modules, this module does not use the output center point and size information to predict the bounding boxes; instead, it uses the information of each of the four boundaries to achieve bounding box prediction. Experimental results on the HRSID dataset show that the proposed method outperforms other deep learning-based instance segmentation methods. Summary of the Invention

[0005] This invention belongs to the field of synthetic aperture radar (SAR) image interpretation technology and discloses a SAR ship instance segmentation method based on a global semantic boundary attention network to address the limited target bounding box localization capability in existing technologies. This method, based on deep learning theory, mainly consists of two parts: a global context information modeling module and a boundary attention prediction module. The global context information modeling module establishes long-range dependencies by repeatedly enhancing the semantic information of features, thereby effectively reducing background interference. The boundary attention prediction module improves the target bounding box localization capability by predicting the target's boundary information twice. Experiments show that on the HRSID dataset, the SAR ship instance segmentation method based on the global semantic boundary attention network achieves an average accuracy (AP) of 57.3%, while the highest average accuracy among other existing deep learning-based SAR ship instance segmentation methods is 55.4%. The SAR ship instance segmentation method based on the global semantic boundary attention network significantly improves the accuracy of ship instance segmentation.

[0006] To facilitate the description of the present invention, the following terms are defined first:

[0007] Definition 1: Traditional HRSID Dataset Acquisition Method

[0008] The HRSID dataset is a commonly used SAR image dataset for ship instance segmentation. Its full name is High-Resolution SAR Images Dataset. This dataset originates from 136 panoramic SAR images with resolutions ranging from 1 meter to 5 meters. Each panoramic SAR image was processed using a sliding window mechanism with a 25% overlap to obtain multiple SAR image slices. Each slice is 800×800 pixels, resulting in a total of 5604 slices containing 16951 ships. 65% of the slices in the HRSID dataset are used as the training set, and the remaining 35% are used as the test set. For details on how to obtain the HRSID dataset, please refer to "Wei S, Zeng X, Qu Q, et al. HRSID: A High-Resolution SAR Images Dataset for Ship Detection and Instance Segmentation[J].IEEE Access,2020,8:1-1."

[0009] Definition 2: Traditional Residual Backbone Network Construction Method

[0010] Residual backbone networks are a commonly used type of backbone network. Proposed by four researchers from Microsoft Research, these convolutional neural networks won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2015 for both image classification and object recognition. Compared to conventional backbone networks, residual backbone networks reduce the probability of vanishing and exploding gradients by adding several residual connections, enabling faster optimization. Therefore, the number of convolutional layers in a residual backbone network can often be more than in a conventional backbone network. A 101-layer residual network is one of the more commonly used structures among residual networks. Specifically, a 101-layer residual network first extracts features using convolutional layers of size 7×7 with 64 kernels and a stride of 2. Then, it uses 3×3 max pooling to achieve a 2x downsampling, which serves as the first-stage feature map output. Finally, by stacking several residual modules, the outputs of the second, third, fourth, and fifth stages are extracted respectively. The difference lies in the number of residual modules used in each stage, and the number of convolutional kernels in the convolutional layers of the residual modules in each stage. Thanks to the cross-connectivity structure in the residual modules, convolutional networks can achieve deep stacking while avoiding gradient vanishing, gradient exploding, and degradation problems. This allows for faster network optimization and the extraction of more discriminative abstract features. For a detailed explanation of classic residual network construction methods, see "K. He et al., 'Deep Residual Learning for Image Recognition,' IEEE Conf.Comput.Vis.Pattern Recognit., 2016, pp.770-778."

[0011] Definition 3: Traditional Regional Recommendation Network Construction Methods

[0012] Region Recommendation Networks (RNRs) were proposed in Faster R-CNN. Addressing the time-consuming nature of the region recommendation algorithm in Fast R-CNN, RNRs replaced the selective search algorithm by introducing the concept of shared convolutional feature maps, thus fusing RNRs and Fast R-CNN into a single network for faster object detection. Furthermore, by pre-setting anchor boxes of different sizes and aspect ratios, RNRs enhance the multi-scale detection capabilities of the detection network, thereby improving object detection accuracy. Specifically, RNRs take an entire image as input and output a series of region recommendation box locations and their confidence scores. The confidence score represents the probability that the region recommendation box is foreground. RNRs consist of two sub-networks. The first sub-network is the backbone network, shared by both RNRs and Fast R-CNN; this mechanism, known as shared convolutional feature maps, aims to conserve computational resources for object detection. The second sub-network comprises an intermediate layer, a classification layer, and a regression layer. The intermediate layer is essentially a fully connected layer; the classification layer and the regression layer are essentially 3×3 convolutional layers. It should be noted that the classification layer and the regression layer are two parallel modules, and their inputs are the outputs of the intermediate layer. In addition, the working principle of the second sub-network is similar to that of the convolution operation, both of which use a sliding window mechanism. Specifically, the input of the second sub-network is a local region of the feature map. During the forward propagation stage, the second sub-network slides on the feature map, and at each position it slides, it calculates the classification and regression information of several anchor boxes corresponding to that position. For the classic method of constructing region recommendation networks, please refer to "Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J].IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6): 1137-1149."

[0013] Definition 4: Construction method of traditional region of interest feature extraction module

[0014] The Region of Interest (ROI) feature extraction module was first proposed in the Fast R-CNN paper. It is used to extract fixed-size local features from the feature map based on the coordinates of the region of interest. The RIO feature extraction module proposed in the Fast R-CNN paper is RoI pooling, whose basic idea is to obtain fixed-size local features through double quantization and max pooling. However, double quantization often leads to some accuracy loss, resulting in inconsistencies between the extracted local features and the coordinates of the region of interest. Therefore, the Mask R-CNN paper proposed RoI Align for RIO feature extraction. Its main idea is to abandon the quantization operation and use bilinear interpolation to calculate the feature values ​​of the corresponding coordinate points, thus ensuring that the extracted local features are consistent with the coordinates of the region of interest. RoI Align has now become the mainstream implementation of the RIO feature extraction module. For detailed information on the construction method of the RIO feature extraction module, please refer to "He K, Gkioxari G, P Dollár, et al. Mask R-CNN[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017."

[0015] Definition 5: Traditional Convolutional Layer Construction Methods

[0016] Convolutional layers are fundamental modules in deep learning neural networks. Their basic function is to extract abstract features from input data, facilitating subsequent tasks such as classification and regression. A convolutional layer typically contains multiple convolutional kernels. Each kernel is a node that weights and sums the values ​​within a small rectangular region of the input feature map or image to produce the output. Each kernel requires manually specified parameters. One type of parameter is the length and width of the node matrix processed by the kernel; the size of this node matrix is ​​also the size of the kernel. Another type of parameter is the depth of the resulting unit node matrix; the depth of the unit node matrix is ​​also the depth of the kernel. During the convolution operation, each kernel slides across the input data, calculating the inner product between the kernel and corresponding positions in the input data. This inner product is then processed by a non-linear function to obtain the final result. Finally, all the results at corresponding positions form a two-dimensional feature map. Each convolutional kernel generates a two-dimensional feature map; the feature maps generated by multiple convolutional kernels are superimposed to form a three-dimensional feature map. Typically, convolutional layers use 3×3 or 5×5 kernels. The depth of the kernel is determined by the number of feature channels in the previous layer, and the number of kernels is determined by the designer. For classic convolutional layer construction methods, please refer to "Fan Lili, Zhao Hongwei, Zhao Haoyu, Hu Huangshui, Wang Zhen. A review of target detection research based on deep convolutional neural networks [J]. Optics and Precision Engineering, 2020, 28(05):1152-1164."

[0017] Definition 6: Traditional pixel reconstruction construction method

[0018] Pixel reassembly was initially proposed for super-resolution image tasks, and has since been applied to image classification and detection tasks. Pixel reassembly is an upsampling method that can effectively enlarge scaled-down feature maps, and can serve as an alternative to deconvolution or nearest-neighbor interpolation.

[0019] Definition 7: Traditional methods for constructing dilated convolutional layers

[0020] Dilated convolutional layers are similar to standard convolutional layers, except that they add a dilation rate parameter, which increases the sampling domain size and ultimately the receptive field. By increasing the receptive field, dilated convolutional layers can extract global information to some extent, enhancing the semantic information in the output feature map and helping the neural network distinguish between the target and background interference.

[0021] Definition 8: Traditional Cascading Operation

[0022] Cascading is an important operation in network architecture design, used to combine features, fuse features extracted by multiple convolutional feature extraction frameworks, or fuse information from the output layer, thereby enhancing the network's feature extraction capabilities.

[0023] Definition 9: Traditional global feature self-attention construction method

[0024] The global feature self-attention module is used to extract non-local features from the input. The basic idea of ​​this module is to calculate the similarity weights of each input pixel with all other pixels, and then use these similarity weights to perform a weighted sum of the corresponding pixels as the output. Compared to convolutional layers, the global feature self-attention module has a larger field of view, not limited to a local field of view, thus it can extract global information and enhance the semantic information in the feature map. For a detailed explanation of classic global feature self-attention construction methods, see "Wang X, Girshick R, Gupta A, et al. Non-local Neural Networks[J].2017."

[0025] Definition 10: Traditional Fully Connected Layer Construction Method

[0026] Fully connected layers are a type of structure in neural networks used for further feature extraction. Unlike convolutional layers, the number of input and output nodes in a fully connected layer needs to be pre-set, and its parameter count and computational cost far exceed those of a convolutional layer. Therefore, fully connected layers often appear only in a specific part of a neural network structure. For a classic example of a fully connected layer method, see "Haoren Wang, Haotian Shi, Ke Lin, Chengjin Qin, Liqun Zhao, Yixiang Huang, Chengliang Liu. A high-precision arrhythmia classification method based on dual fully connected neural network[J]. Biomedical Signal Processing and Control, 2020, 58".

[0027] Definition 11: Traditional Convolutional Attention Module Construction Method

[0028] The convolutional attention module mainly consists of three parts: pooling, convolution, and activation function. Specifically, for the input feature F, it first undergoes a channel-level max pooling and a channel-level average pooling to obtain two two-dimensional feature maps. These two feature maps are then concatenated, and finally, a spatial attention map M is obtained through a convolutional kernel activation function. S The specific expression for the spatial attention module is shown in equation M. S (F)=σ(f 7×7 As shown in ([AvgPool(F); MaxPool(F)])), where f 7×7The σ represents the convolution operation with a kernel size of 7×7, and σ represents the sigmoid activation function. It's important to note that in this chapter, to extract features from multiple receptive fields, two spatial attention modules are used in parallel, with convolutional kernel sizes of 7×7 and 3×3 respectively. A classic method for constructing convolutional attention modules is illustrated in "Woo, S.; Park, J.; Lee, JY; Kweon, ISJS, Cham, CBAM: Convolutional Block Attention Module. 2018."

[0029] Definition 12: Traditional Boundary Coarse-Localization Subnetwork Construction Method

[0030] The coarse boundary localization subnetwork takes boundary features as input and outputs the approximate location of the corresponding boundary. Specifically, the coarse boundary localization subnetwork divides the target space into multiple discrete intervals. For a given boundary feature, the coarse boundary localization subnetwork only provides the parameter (i.e., s) indicating which interval the corresponding boundary belongs to. x-right ,s x-left ,s y-right , and s y-left ), without providing more precise boundary regression values. Wherein, s x-right s represents the confidence level of the boundary on the right side of the vertical direction. x-left s represents the confidence level of the boundary on the left side of the vertical direction. y-right s represents the right-hand side of the horizontal boundary confidence level. y-left This represents the confidence level of the horizontal boundary on the left side. For details on the construction method of the classic boundary coarse localization sub-network, please refer to "Wang J, Zhang W, Cao Y, et al. Side-Aware Boundary Localization for More Precise Object Detection[J].2019.".

[0031] Definition 13: Traditional method for constructing boundary-precise localization subnetworks

[0032] The boundary fine localization subnetwork further refines the boundary position based on the boundary coarse localization. Similar to traditional boundary classification and regression, this process uses the predicted bounding boxes from the coarse localization as prior knowledge, outputting the coordinate and size offsets between the target box and the predicted box, thus obtaining more accurate boundary predictions. For a detailed explanation of the classic boundary fine localization subnetwork construction method, please refer to "Wang J, Zhang W, Cao Y, et al. Side-Aware Boundary Localization for More Precise Object Detection[J].2019."

[0033] Definition 14: Traditional Masked Subnetwork Construction Method

[0034] The mask subnetwork is extracted in Mask R-CNN. This subnetwork takes the result of the boundary prediction network as input and outputs a pixel-level binary classification result of the target region, which can achieve pixel-level differentiation between the target and the background, thereby extracting the edge information of the target. For the classic mask subnetwork construction method, please refer to "He K, Gkioxari G, P Dollár, et al. Mask R-CNN[J].IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017.".

[0035] Definition 15: The classic Adam algorithm

[0036] The classic Adam algorithm is an extension of stochastic gradient descent and has recently been widely used in deep learning applications in computer vision and natural language processing. Classic Adam differs from classic stochastic gradient descent. Stochastic gradient descent maintains a single learning rate for all weight updates, and this learning rate remains unchanged during training. Each network weight, however, maintains a learning rate that is adjusted individually as learning progresses. This method computes adaptive learning rates for different parameters from the budgets of the first and second moments of the gradient. For a detailed explanation of the classic Adam algorithm, see "Kingma, D.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2014, arXiv:1412.6980."

[0037] Definition 16: Traditional forward propagation method

[0038] Forward propagation is one of the most fundamental methods in deep learning. It mainly involves forward reasoning based on the parameters and connection methods in the network to obtain the network's output.

[0039] Definition 17: Traditional Nonmaximum Suppression Methods

[0040] Non-maximum suppression (NMS) is an algorithm used in object detection to remove redundant bounding boxes. In the forward propagation results of classic detection networks, multiple bounding boxes often correspond to the same object. Therefore, an algorithm is needed to select the highest-quality bounding box with the highest score from these multiple boxes. Non-maximum suppression performs a local maximum search by calculating an overlap threshold.

[0041] Definition 18: Traditional methods for calculating recall and precision

[0042] Recall R refers to the number of correctly predicted samples out of all positive samples, expressed as: Precision P refers to the proportion of correct predictions out of the total number of predictions that are positive. The expression is: Wherein, TP (true positive) represents a positive sample predicted as positive by the model; FN (false negative) represents a negative sample predicted as negative by the model; and FP (false positive) represents a positive sample predicted as negative by the model. The recall and precision curves P(R) refer to functions with R as the independent variable and P as the dependent variable. For details on how to calculate the values ​​of the above parameters, please refer to the reference "Li Hang. Statistical Learning Methods [M]. Beijing: Tsinghua University Press, 2012."

[0043] This invention provides a SAR ship instance segmentation method based on a global semantic boundary attention network, which includes the following steps:

[0044] Step 1: Initialize the dataset

[0045] The HRSID dataset is obtained using the traditional method described in Definition 1, and the training set in the HRSID dataset is denoted as D. test The training set is denoted as D. train .

[0046] Step 2: Build a forward propagation network

[0047] Step 2.1: Build the ResNet-101 backbone network

[0048] A residual network with 101 layers, denoted as Res-101, is constructed using the traditional residual backbone network construction method in Definition 2.

[0049] Step 2.2: Build a regional recommendation network

[0050] The classic regional recommendation network construction method in Definition 3 is used to construct the regional recommendation network. The ResNet-101 backbone network Res-101 obtained in step 2.1 is used as a sub-network in the regional recommendation network, and the constructed regional recommendation network is denoted as RPN0.

[0051] Step 2.3: Build the feature extraction module

[0052] The feature extraction module is constructed using the traditional region of interest feature extraction module construction method in Definition 4, and the constructed feature extraction module is denoted as FExtract.

[0053] Step 2.4: Build the global context information modeling module

[0054] First, two convolutional layers, denoted as conv1 and conv2, are constructed using the traditional convolutional layer construction method in Definition 5. Then, a pixel reorganization module, denoted as pixelshuffle, is constructed using the traditional pixel reorganization construction method in Definition 6. According to the expression... Define a Softmax layer, denoted as softmax0, where z i Let C represent the feature value of the i-th node in the input feature map, and C represent the number of channels in the input feature map. Concatenate conv1, conv2, pixelshuffle, and softmax, denoted as kplayer. According to the expression... Construct a feature recombination layer, denoted as cz. Wherein, (i,j) represents l, F (i+n,j+m) W represents the eigenvector in F located at (i+n,j+m). l'(n,m) Represents W l' The weights located at (n, m) are used. Combining kplayer and czlayer completes the construction of the content-aware feature reorganization subnetwork, which is denoted as carafe.

[0055] Then, using the traditional dilated convolutional layer construction method in Definition 7, dilated convolutional layers with dilation rates of 2, 3, 4, and 5 are constructed, denoted as d1, d2, d3, and d4, respectively. A concatenated module, denoted as concate, is constructed using the traditional concatenation operation construction method in Definition 8. A convolutional layer, denoted as conv3, is constructed using the traditional convolutional layer construction method in Definition 6. After concatenating d1, d2, d3, and d4 in parallel, they are then concatenated with concate and con3 in sequence, thus completing the construction of the multi-view feature extraction subnetwork. This constructed multi-view feature extraction subnetwork is denoted as mrblock.

[0056] Finally, the traditional global feature self-attention construction method in Definition 9 is used to construct a global feature self-attention subnetwork, which is denoted as sablock.

[0057] The feature recombination subnetwork carafe, the multi-view feature extraction subnetwork mrblock, and the global feature self-attention subnetwork sablock are concatenated in sequence to obtain the global context information modeling module, denoted as GCB.

[0058] Step 2.5: Build the boundary attention prediction module

[0059] Based on the traditional fully connected layer construction method defined in Definition 10, three fully connected layers are constructed, denoted as fc1, fc2, and fc3 respectively. Concatenating fc1, fc2, and fc3 completes the construction of the classification branch, denoted as CLBranch, and the classification result output by CLBranch is denoted as s.

[0060] A convolutional attention module, denoted as CBAM, is constructed using the traditional convolutional attention module construction method defined in Definition 11. Four convolutional layers, denoted as conv4, conv5, conv6, and conv7, are constructed using the traditional convolutional layer construction method defined in Definition 5. According to the expression... Define two softmax layers, denoted as softmax1 and softmax2, where z i Let C be the feature value of the i-th node in the input feature map, and C represent the number of channels in the input feature map. Concatenate conv4, softmax1, and conv5 in sequence, denoted as branchx. Concatenate conv6, softmax2, and conv7 in sequence, denoted as branchchy. Add the parallel concatenation of branchx and branchchy after the CBAM module to complete the boundary attention prediction module, denoted as baff.

[0061] The boundary coarse localization subnetwork is constructed using the traditional boundary coarse localization subnetwork construction method defined in Definition 12. The constructed boundary coarse localization subnetwork is denoted as bbcl, and the four outputs of bbcl are denoted as s. x-right ,s x-left ,s y-right ands y-left .

[0062] The traditional boundary fine-localization subnetwork construction method defined in Definition 13 is used to construct the boundary fine-localization subnetwork, and the constructed boundary fine-localization subnetwork is denoted as brfl.

[0063] The classification result s output by CLBranch in step 2.5 and the classification result s output by bbcl in step 2.6 are used. x-right ,s x-left ,s y-right ,and s y-left As input, according to the formula The calculation is performed, which completes the construction of the boundary-guided classification re-scoring sub-network, denoted as cbcr.

[0064] By concatenating baff, bbcl, brfl, and cbcr in sequence, the boundary attention prediction module is completed, denoted as BABP.

[0065] Step 2.6: Construct the mask subnet

[0066] According to Definition 14, the traditional mask subnetwork construction method is used to construct a mask subnetwork, and the constructed mask subnetwork is denoted as MASK.

[0067] Step 2.7: Build an instance segmentation cascaded network

[0068] The feature extraction module FExtract obtained in step 2.3, the global context information modeling module GCB obtained in step 2.4, the boundary attention prediction module BABP obtained in step 2.5, and the mask sub-network MASK obtained in step 2.6 are concatenated to obtain the first instance segmentation network, denoted as SEG1.

[0069] Repeat steps 2.3, 2.4, 2.5, and 2.6, and concatenate the modules or sub-networks obtained in each step to obtain the second instance segmentation network, denoted as SEG2.

[0070] Repeat steps 2.3, 2.4, 2.5, and 2.6, and concatenate the modules or sub-networks obtained in each step to obtain the third instance segmentation network, denoted as SEG3.

[0071] The first instance segmentation network SEG1, the second instance segmentation network SEG2, and the third instance segmentation network SEG3 are connected in series to form the instance segmentation cascade network, denoted as CASEG0.

[0072] Step 3: Training the Region Recommendation Network

[0073] Set the iteration parameter epoch, and initialize the epoch value to 1.

[0074] Step 3.1: Perform forward propagation on the regional recommendation network.

[0075] The training set D obtained in step 1 train As input to the region recommendation network RPN0, the training set D is processed according to the forward propagation method in Definition 16. train The data is fed into the regional recommendation network RPN0 for computation, and the output of network RPN0 is denoted as Result0.

[0076] Step 3.2: Sample the forward propagation results.

[0077] Combine the input Result0 obtained in step 3.1 with the training set D train As input, according to the formula The calculation method involves calculating the IOU value of each recommendation box in Result0. Outputs with an IOU greater than 0.5 in Result0 are designated as positive samples, denoted as Result0p; outputs with an IOU less than 0.5 are designated as negative samples, denoted as Result0n. The total number of negative samples in Result0n is M. The required number of negative samples is manually input, denoted as N; the required number of intervals for equally spaced IOUs is manually input, denoted as n. b Let M be the number of samples in the i-th IOU interval. i Set the random sampling probability of the i-th interval to be... For each IOU interval, random sampling is performed, and the sampling results of all IOU intervals of the negative samples are denoted as Result0ns.

[0078] Count the number of positive samples in Result0p, denoted as P. Set the random sampling probability to... Randomly sample Result0p and record the positive sample results as Result0ps.

[0079] Step 3.3: Train and optimize the region recommendation network.

[0080] Using the positive sample sampling results Result0ps and negative sample sampling results Result0ns obtained in step 3.2 as input, the region recommendation network is trained and optimized according to the classic Adam algorithm in Definition 15. The trained and optimized region recommendation network RPN1 is obtained.

[0081] Step 4: Train the instance segmentation cascaded network

[0082] Step 4.1: Perform forward propagation on the instance segmentation concatenation network.

[0083] The training set D obtained in step 1 train As input to the instance segmentation cascaded network CASEG0, the training set D is processed according to the traditional forward propagation method in Definition 16. train The data is fed into the instance segmentation cascade network CASEG0 for computation, and the output of the instance segmentation cascade network CASEG0 is denoted as Result1.

[0084] Step 4.2: Train and optimize the instance segmentation concatenated network.

[0085] Using the output Result1 of the instance segmentation concatenated network CASEG0 obtained in step 4.1 as input, the instance segmentation concatenated network is trained and optimized according to the classic Adam algorithm in Definition 15. The trained and optimized instance segmentation concatenated network CASEG1 is then obtained.

[0086] Step 5: Perform alternating training

[0087] Determine if the epoch set in step 3 is equal to 12. If the epoch is not equal to 12, set epoch = epoch + 1, RPN0 = RPN1, and CASEG0 = CASEG1, and repeat steps 3.1, 3.2, 3.3, 4.1, and 4.2 in sequence, then return to step 5 to check the epoch again; if the epoch is equal to 12, then denote the trained region recommendation network RPN1 and the trained instance segmentation concatenated network CASEG1 as network GCBAN, and then proceed to step 6.

[0088] Step 6, Evaluation Method

[0089] Step 6.1, Forward Propagation

[0090] Using the network GCBAN obtained in step 6 and the test set D obtained in step 1 test As input, the traditional forward propagation method defined in Definition 16 is used to obtain the detection result, denoted as R.

[0091] Using the detection result R as input, the traditional nonmaximum suppression method in Definition 17 is used to remove redundant boxes in the detection result R1. The specific steps are as follows:

[0092] Step (1) First, let the highest-scoring bounding box in the detection result R1 be denoted as BS;

[0093] Step (2) Then the calculation formula is as follows: Calculate the overlap ratio threshold (IoU) of all boxes in the detection result R1; discard boxes with IoU > 0.5;

[0094] Step (3) Select the highest-scoring box (BS) from the remaining boxes;

[0095] Repeat the process of calculating IoU and discarding boxes in step (2) above until there are no more boxes to discard. The remaining boxes are the final detection results, denoted as R. F .

[0096] Step 6.2, Calculate the indicators

[0097] The detection result R obtained in step 6.1 F As input, the traditional recall and precision calculation methods in Definition 18 are used to obtain the network's precision P, recall R, and precision-recall curve P(R); the formula is used... Calculate the SAR ship instance segmentation accuracy metrics AP and AP based on equilibrium learning. 50 AP 75 AP S AP M APL .

[0098] The innovation of this invention lies in the introduction of a global context information modeling module and a boundary attention prediction module, thereby solving the problem of limited target box localization capability in existing deep learning-based SAR ship instance segmentation methods. The AP (Achievement Probability) for SAR image ship instance segmentation using this method is 57.3%, exceeding the suboptimal SAR image ship instance segmentation method by 1.9 percentage points. 50 The accuracy rate was 88.6%, exceeding the suboptimal SAR image ship instance segmentation method by 2.8 percentage points; the AP of SAR image ship instance segmentation using this method was [missing information]. 75 The accuracy rate was 68.9%, exceeding the suboptimal SAR image ship instance segmentation method by 2.0 percentage points; the AP of SAR image ship instance segmentation using this method was [missing information]. S The success rate was 57%, exceeding the suboptimal SAR image ship instance segmentation method by 2.1 percentage points; the AP of SAR image ship instance segmentation using this method was [missing information]. M The accuracy rate was 64.3%, exceeding the suboptimal SAR image ship instance segmentation method by 0.8 percentage points; the AP of SAR image ship instance segmentation using this method was [missing information]. L The accuracy rate is 25.9%, exceeding the suboptimal SAR image ship instance segmentation method by 6.2 percentage points. In summary, this method can achieve good target bounding box localization and has excellent SAR ship instance segmentation accuracy.

[0099] The advantage of this invention is that it can overcome the problem of limited target box localization capability in the prior art and improve the instance segmentation accuracy of ships in SAR images. Attached Figure Description

[0100] Figure 1 This is a flowchart illustrating the SAR image ship instance segmentation method based on a global semantic boundary attention network in this invention.

[0101] Wherein, 1 represents the boundary attention prediction module, and 2 represents the global top-bottom information modeling module;

[0102] Figure 2 This refers to the instance segmentation accuracy index of the SAR image ship instance segmentation method based on global semantic boundary attention network in this invention. Detailed Implementation

[0103] The following is in conjunction with the appendix Figure 1 Appendix Figure 2 The present invention will be described in further detail below.

[0104] Step 1: Initialize the dataset

[0105] Obtain the HRSID dataset according to the method described in Definition 1, and denote the test set in the HRSID dataset as D. test The training set is denoted as D. train .

[0106] Step 2: Build a forward propagation network

[0107] Step 2.1: Build the ResNet-101 backbone network

[0108] like Figure 1 As shown, a residual network with 101 layers, denoted as Res-101, is constructed using the classic residual backbone network construction method in Definition 2.

[0109] Step 2.2: Build a regional recommendation network

[0110] like Figure 1 As shown, the classic regional recommendation network construction method in Definition 3 is used to construct the regional recommendation network. The ResNet-101 backbone network Res-101 obtained in step 2.1 is used as a sub-network in the regional recommendation network, and the constructed regional recommendation network is denoted as RPN0.

[0111] Step 2.3: Build the feature extraction module

[0112] like Figure 1 As shown, the feature extraction module is constructed using the region of interest feature extraction module construction method in Definition 4, and the constructed feature extraction module is denoted as FExtract.

[0113] Step 2.4: Build the global context information modeling module

[0114] First, two convolutional layers, denoted as conv1 and conv2, are constructed using the convolutional layer construction method in Definition 5. Then, a pixel reorganization module, denoted as pixelshuffle, is constructed using the pixel reorganization construction method in Definition 6. According to the expression... Define a Softmax layer, denoted as softmax0, where z i Let C represent the feature value of the i-th node in the input feature map, and C represent the number of channels in the input feature map. Concatenate conv1, conv2, pixelshuffle, and softmax, denoted as kplayer. According to the expression... Construct a feature recombination layer, denoted as cz. Wherein, (i,j) represents l, F (i+n,j+m) W represents the eigenvector in F located at (i+n,j+m). l'(n,m) Represents W l'The weights located at (n, m) are used. Combining kplayer and czlayer completes the construction of the content-aware feature reorganization subnetwork, which is denoted as carafe.

[0115] Then, using the dilated convolutional layer construction method in Definition 7, dilated convolutional layers with dilation rates of 2, 3, 4, and 5 are constructed, denoted as d1, d2, d3, and d4, respectively. A concatenated module, denoted as concate, is constructed using the concatenation operation construction method in Definition 8. A convolutional layer, denoted as conv3, is constructed using the convolutional layer construction method in Definition 6. After concatenating d1, d2, d3, and d4 in parallel, they are then concatenated with concate and con3 in sequence, thus completing the construction of the multi-view feature extraction subnetwork. This constructed multi-view feature extraction subnetwork is denoted as mrblock.

[0116] Finally, the global feature self-attention construction method in Definition 9 is used to construct the global feature self-attention sub-network, and the constructed global feature self-attention sub-network is denoted as sablock.

[0117] like Figure 1 As shown, the feature reorganization subnetwork carafe, the multi-view feature extraction subnetwork mrblock, and the global feature self-attention subnetwork sablock are concatenated in sequence to obtain the global context information modeling module, denoted as GCB.

[0118] Step 2.5: Build the boundary attention prediction module

[0119] Three fully connected layers, denoted as fc1, fc2, and fc3, are constructed according to the fully connected layer construction method in Definition 10. Concatenating fc1, fc2, and fc3 completes the construction of the classification branch, denoted as CLBranch, and the classification result output by CLBranch is denoted as s.

[0120] A convolutional attention module, denoted as CBAM, is constructed using the convolutional attention module construction method defined in Definition 11. Four convolutional layers, denoted as conv4, conv5, conv6, and conv7, are constructed using the convolutional layer construction method defined in Definition 5. According to the expression... Define two softmax layers, denoted as softmax1 and softmax2, where z iLet C be the feature value of the i-th node in the input feature map, and C represent the number of channels in the input feature map. Concatenate conv4, softmax1, and conv5 in sequence, denoted as branchx. Concatenate conv6, softmax2, and conv7 in sequence, denoted as branchchy. Add the parallel concatenation of branchx and branchchy after the CBAM module to complete the boundary attention prediction module, denoted as baff.

[0121] The boundary coarse localization subnetwork is constructed using the boundary coarse localization subnetwork construction method defined in Definition 12. The constructed boundary coarse localization subnetwork is denoted as bbcl, and the four outputs of bbcl are denoted as s. x-right ,s x-left ,s y-right ,and s y-left .

[0122] The boundary fine-localization subnetwork is constructed using the boundary fine-localization subnetwork construction method of Definition 13, and the constructed boundary fine-localization subnetwork is denoted as brfl.

[0123] The classification result s output by CLBranch in step 2.5 and the classification result s output by bbcl in step 2.6 are used. x-right ,s x-left ,s y-right ,and s y-left As input, according to the formula The calculation is performed, which completes the construction of the boundary-guided classification re-scoring sub-network, denoted as cbcr.

[0124] like Figure 1 As shown, by connecting baff, bbcl, brfl, and cbcr in sequence, the boundary attention prediction module is completed, denoted as BABP.

[0125] Step 2.6: Construct the mask subnet

[0126] like Figure 1 As shown, a mask subnetwork is constructed according to the mask subnetwork construction method in Definition 14, and the constructed mask subnetwork is denoted as MASK.

[0127] Step 2.7: Build an instance segmentation cascaded network

[0128] The feature extraction module FExtract obtained in step 2.3, the global context information modeling module GCB obtained in step 2.4, the boundary attention prediction module BABP obtained in step 2.5, and the mask sub-network MASK obtained in step 2.6 are concatenated to obtain the first instance segmentation network, denoted as SEG1.

[0129] Repeat steps 2.3, 2.4, 2.5, and 2.6, and concatenate the modules or sub-networks obtained in each step to obtain the second instance segmentation network, denoted as SEG2.

[0130] Repeat steps 2.3, 2.4, 2.5, and 2.6, and concatenate the modules or sub-networks obtained in each step to obtain the third instance segmentation network, denoted as SEG3.

[0131] The first instance segmentation network SEG1, the second instance segmentation network SEG2, and the third instance segmentation network SEG3 are connected in series to form the instance segmentation cascade network, denoted as CASEG0.

[0132] Step 3: Training the Region Recommendation Network

[0133] Set the iteration parameter epoch, and initialize the epoch value to 1.

[0134] Step 3.1: Perform forward propagation on the regional recommendation network.

[0135] The training set D obtained in step 1 train As input to the region recommendation network RPN0, the training set D is processed according to the forward propagation method in Definition 16. train The data is fed into the regional recommendation network RPN0 for computation, and the output of network RPN0 is recorded as Result0.

[0136] Step 3.2: Sample the forward propagation results.

[0137] Combine the input Result0 obtained in step 3.1 with the training set D train As input, according to the formula The calculation method involves calculating the IOU value of each recommendation box in Result0. Outputs with an IOU greater than 0.5 in Result0 are designated as positive samples, denoted as Result0p; outputs with an IOU less than 0.5 are designated as negative samples, denoted as Result0n. The total number of negative samples in Result0n is M. The required number of negative samples is manually input, denoted as N; the required number of intervals for equally spaced IOUs is manually input, denoted as n. b Let M be the number of samples in the i-th IOU interval. i Set the random sampling probability of the i-th interval to be... For each IOU interval, random sampling is performed, and the sampling results of all IOU intervals of the negative samples are denoted as Result0ns.

[0138] Count the number of positive samples in Result0p, denoted as P. Set the random sampling probability to... Randomly sample Result0p and record the positive sample results as Result0ps.

[0139] Step 3.3: Train and optimize the region recommendation network.

[0140] Using the positive sample sampling results Result0ps and negative sample sampling results Result0ns obtained in step 3.2 as input, the region recommendation network is trained and optimized according to the classic Adam algorithm in Definition 15. The trained and optimized region recommendation network RPN1 is obtained.

[0141] Step 4: Train the instance segmentation cascaded network

[0142] Step 4.1: Perform forward propagation on the instance segmentation concatenation network.

[0143] The training set D obtained in step 1 train As input to the instance segmentation cascaded network CASEG0, the training set D is processed according to the forward propagation method in Definition 16. train The data is fed into the instance segmentation cascade network CASEG0 for computation, and the output of the instance segmentation cascade network CASEG0 is recorded as Result1.

[0144] Step 4.2: Train and optimize the instance segmentation concatenated network.

[0145] Using the output Result1 of the instance segmentation concatenated network CASEG0 obtained in step 4.1 as input, the instance segmentation concatenated network is trained and optimized according to the classic Adam algorithm in Definition 15. The trained and optimized instance segmentation concatenated network CASEG1 is then obtained.

[0146] Step 5: Perform alternating training

[0147] Determine if the epoch set in step 3 is equal to 12. If the epoch is not equal to 12, set epoch = epoch + 1, RPN0 = RPN1, and CASEG0 = CASEG1, and repeat steps 3.1, 3.2, 3.3, 4.1, and 4.2 in sequence, then return to step 5 to check the epoch again; if the epoch is equal to 12, then denote the trained region recommendation network RPN1 and the trained instance segmentation concatenated network CASEG1 as network GCBAN, and then proceed to step 6.

[0148] Step 6, Evaluation Method

[0149] Step 6.1, Forward Propagation

[0150] Using the network GCBAN obtained in step 6 and the test set D obtained in step 1test As input, the traditional forward propagation method defined in Definition 16 is used to obtain the detection result, denoted as R.

[0151] Using the detection result R as input, the traditional nonmaximum suppression method in Definition 17 is used to remove redundant boxes in the detection result R1. The specific steps are as follows:

[0152] Step (1) First, let the highest-scoring bounding box in the detection result R1 be denoted as BS;

[0153] Step (2) Then the calculation formula is as follows: Calculate the overlap ratio threshold (IoU) of all boxes in the detection result R1; discard boxes with IoU > 0.5;

[0154] Step (3) Select the highest-scoring box (BS) from the remaining boxes;

[0155] Repeat the process of calculating IoU and discarding boxes in step (2) above until there are no more boxes to discard. The remaining boxes are the final detection results, denoted as R. F .

[0156] Step 6.2, Calculate the indicators

[0157] The detection result R obtained in step 6.1 F As input, the traditional recall and precision calculation methods in Definition 18 are used to obtain the network's precision P, recall R, and precision-recall curve P(R); the formula is used... Calculate the average accuracy mAP for SAR ship instance segmentation based on balanced learning.

Claims

1. A SAR ship instance segmentation method based on a global semantic boundary attention network, characterized by its... Includes the following steps: Step 1: Initialize the dataset The HRSID dataset is obtained using the traditional method, and the test set in the HRSID dataset is denoted as D. test The training set is denoted as D. train ; Step 2: Build a forward propagation network Step 2.1: Build the ResNet-101 backbone network A residual network with 101 layers, denoted as Res-101, is constructed using the classic residual backbone network construction method. Step 2.2: Build a regional recommendation network The region recommendation network is constructed using the classic region recommendation network construction method. The ResNet-101 backbone network Res-101 obtained in step 2.1 is used as a sub-network in the region recommendation network, and the constructed region recommendation network is denoted as RPN0. Step 2.3: Build the feature extraction module A feature extraction module is constructed using the traditional method for constructing region of interest feature extraction modules. The constructed feature extraction module is denoted as FExtract. Step 2.4: Build the global context information modeling module First, two convolutional layers are constructed using the traditional convolutional layer construction method, denoted as conv1 and conv2 respectively. Then, a pixel reorganization module, denoted as pixelshuffle, is constructed using the traditional pixel reorganization construction method. According to the expression Define a Softmax layer, denoted as softmax0, where z i Let be the feature value of the i-th node in the input feature map, and C represent the number of channels in the input feature map; conv1, conv2, pixelshuffle, and softmax are concatenated and denoted as kplayer; According to the expression Construct a feature recombination layer, denoted as cz; where, (i,j) represents l, F (i+n,j+m) W represents the eigenvector in F located at (i+n,j+m). l'(n,m) Represents W l' The weights located at (n,m) are used; kplayer and czlayer are combined to complete the construction of the content-aware feature recombination subnetwork, and the constructed content-aware feature recombination subnetwork is denoted as carafe; Then, dilated convolutional layers with dilation rates of 2, 3, 4, and 5 are constructed using the traditional dilated convolutional layer construction method, denoted as d1, d2, d3, and d4 respectively; a concatenated module is constructed using the traditional concatenation operation construction method, denoted as concate; a convolutional layer is constructed using the traditional convolutional layer construction method, denoted as conv3; d1, d2, d3, and d4 are concatenated in parallel and then concatenated with concate and conv3 in sequence, thus completing the construction of the multi-view feature extraction sub-network, which is denoted as mrblock; Finally, the global feature self-attention sub-network is constructed using the traditional global feature self-attention construction method, and the constructed global feature self-attention sub-network is denoted as sablock; The feature recombination subnetwork carafe, the multi-view feature extraction subnetwork mrblock, and the global feature self-attention subnetwork sablock are concatenated in sequence to obtain the global context information modeling module, denoted as GCB; Step 2.5: Build the boundary attention prediction module Three fully connected layers are constructed using the traditional fully connected layer construction method, denoted as fc1, fc2, and fc3 respectively. Fc1, fc2, and fc3 are concatenated to complete the construction of the classification branch, denoted as CLBranch. The classification result output by CLBranch is denoted as s. A convolutional attention module, denoted as CBAM, is constructed using the traditional method for constructing convolutional attention modules. Four convolutional layers, denoted as conv4, conv5, conv6, and conv7, are constructed using the traditional method for constructing convolutional layers. According to the expression Define two softmax layers, denoted as softmax1 and softmax2, where z i Let C be the feature value of the i-th node in the input feature map, and C represent the number of channels in the input feature map. Concatenate conv4, softmax1, and conv5 in order, and denote it as branchx. Concatenate conv6, softmax2, and conv7 in order, and denote it as branchy. Add the parallel concatenation of branchx and branchy after the CBAM module to complete the construction of the boundary attention prediction module, and denote it as baff. A coarse boundary localization subnetwork is constructed using the traditional method. The constructed coarse boundary localization subnetwork is denoted as bbcl, and its four outputs are denoted as S. x-right ,S x-left ,S y-right , and S y-left ; A boundary fine-localization subnetwork is constructed using the traditional boundary fine-localization subnetwork construction method, and the constructed boundary fine-localization subnetwork is denoted as brfl; The classification result s output by CLBranch in step 2.5 and the classification result S output by bbcl in step 2.6 are used. x-right ,S x-left ,S y-right , and S y-left As input, according to the formula The calculation is performed, which completes the construction of the boundary-guided classification re-scoring sub-network, denoted as cbcr; By concatenating baff, bbcl, brfl, and cbcr in sequence, the boundary attention prediction module is completed, denoted as BABP. Step 2.6: Construct the mask subnet A mask subnetwork is constructed according to the traditional mask subnetwork construction method, and the constructed mask subnetwork is denoted as MASK. Step 2.7: Build an instance segmentation cascaded network The feature extraction module FExtract obtained in step 2.3, the global context information modeling module GCB obtained in step 2.4, the boundary attention prediction module BABP obtained in step 2.5, and the mask sub-network MASK obtained in step 2.6 are concatenated to obtain the first instance segmentation network, denoted as SEG1. Repeat steps 2.3, 2.4, 2.5, and 2.6, and concatenate the modules or sub-networks obtained in each step to obtain the second instance segmentation network, denoted as SEG2; Repeat steps 2.3, 2.4, 2.5, and 2.6, and concatenate the modules or sub-networks obtained in each step to obtain the third instance segmentation network, denoted as SEG3; The first instance segmentation network SEG1, the second instance segmentation network SEG2, and the third instance segmentation network SEG3 are connected in series to form the instance segmentation cascaded network, denoted as CASEG0. Step 3: Training the Region Recommendation Network Set the iteration parameter epoch, and initialize the epoch value to 1; Step 3.1: Perform forward propagation on the regional recommendation network. The training set D obtained in step 1 train As input to the region recommendation network RPN0, the training set D is processed using the forward propagation method. train The data is fed into the regional recommendation network RPN0 for computation, and the output of network RPN0 is denoted as Result0. Step 3.2: Sample the forward propagation results. Combine the input Result0 obtained in step 3.1 with the training set D train As input, according to the formula The calculation method involves calculating the IOU value of each recommendation box in Result0. Outputs with an IOU greater than 0.5 in Result0 are designated as positive samples, denoted as Result0p; outputs with an IOU less than 0.5 in Result0 are designated as negative samples, denoted as Result0n; the total number of negative samples in Result0n is M; the required number of negative samples is manually input, denoted as N; and the required number of intervals for equally divided IOUs is manually input, denoted as n. b Let M be the number of samples in the i-th IOU interval. i Set the random sampling probability of the i-th interval to be... For each IOU interval, random sampling is performed, and the sampling results of all IOU intervals of the negative samples are denoted as Result0ns; Count the number of positive samples in Result0p, denoted as P; set the random sampling probability to... Randomly sample Result0p and record the positive sample sampling results as Result0ps; Step 3.3: Train and optimize the region recommendation network. Using the positive sample sampling result Result0ps and negative sample sampling result Result0ns obtained in step 3.2 as input, the region recommendation network is trained and optimized according to the classic Adam algorithm; the trained and optimized region recommendation network RPN1 is obtained. Step 4: Train the instance segmentation cascaded network Step 4.1: Perform forward propagation on the instance segmentation concatenation network. The training set D obtained in step 1 train As input to the instance segmentation cascaded network CASEG0, the training set D is processed using the traditional forward propagation method. train The data is fed into the instance segmentation concatenation network CASEG0 for computation, and the output of the instance segmentation concatenation network CASEG0 is denoted as Result1. Step 4.2: Train and optimize the instance segmentation concatenated network. Using the output Result1 of the instance segmentation concatenated network CASEG0 obtained in step 4.1 as input, the instance segmentation concatenated network is trained and optimized according to the classic Adam algorithm; thus, the trained and optimized instance segmentation concatenated network CASEG1 is obtained. Step 5: Perform alternating training Determine if the epoch set in step 3 is equal to 12; if the epoch is not equal to 12, then set epoch = epoch + 1, RPN0 = RPN1, CASEG0 = CASEG1, and repeat steps 3.1, 3.2, 3.3, 4.1, and 4.2 in sequence, and then return to step 5 to determine the epoch again; if the epoch is equal to 12, then denote the trained region recommendation network RPN1 and the trained instance segmentation concatenated network CASEG1 as network GCBAN, and then proceed to step 6; Step 6, Evaluation Method Step 6.1, Forward Propagation Using the network GCBAN obtained in step 6 and the test set D obtained in step 1 test As input, the traditional forward propagation method is used to obtain the detection result, denoted as R; Using the detection result R as input, the traditional non-maximum suppression method is used to remove redundant boxes in the detection result R1. The specific steps are as follows: Step (1) First, let the highest-scoring bounding box in the detection result R1 be denoted as BS; Step (2) Then the calculation formula is as follows: Calculate the overlap threshold IoU for all boxes in the detection result R1; discard boxes with IoU > 0.5; Step (3) Select the highest-scoring box (BS) from the remaining boxes; Repeat the process of calculating IoU and discarding boxes in step (2) above until there are no more boxes to discard. The remaining boxes are the final detection results, denoted as R. F ; Step 6.2, Calculate the indicators The detection result R obtained in step 6.1 F As input, using traditional recall and precision calculation methods, the network's precision P, recall R, and precision-recall curve P(R) are obtained; the formula is used... Calculate the SAR ship instance segmentation accuracy metrics AP and AP based on equilibrium learning. 50 AP 75 AP S AP M AP L .

Images