A real-time coastline extraction method

Abstract

The present application provides a real-time coastline extraction method, which belongs to the field of scene perception and learning technology, and specifically includes the following steps: step 1, based on existing coastline annotation; step 2, training and test data generation, the training data includes coastline annotation map and coastline unlabeled map, and the test data includes coastline annotation map and coastline unlabeled map; step 3, deep learning network construction and training based on semantic segmentation; step 4, using the coastline extraction network to realize real-time coastline extraction of test data. Through the processing scheme of the present application, real-time coastline extraction is realized in complex land and sea scenes.

Description

Technical Field

[0001] The present application relates to the field of scene perception and learning, and in particular to a real-time coastline extraction method. Background Art

[0002] The existing coastline extraction technology is achieved by segmenting the land and sea into regions, and using the threshold segmentation method to perform rough regional segmentation on the radar echo data. There is also a method of using land and sea electronic maps to achieve coastline extraction, which is to match the map prior information with the radar echo data, and use the matching results to perform regional annotation on the radar echo data.

[0003] The threshold segmentation method is used to perform rough regional segmentation on the radar echo data. The threshold segmentation method that relies solely on the echo intensity is prone to misjudge large and medium-sized targets with strong echoes docked on the sea as land information; at the same time, it is difficult to obtain accurate land and sea classification results for geographical features such as concave sea surfaces when segmenting only by the intensity threshold; regional segmentation is achieved by using the method of registering electronic map prior information and radar data. Since the map update has a certain period, there is a problem of low real-time performance. Some artificial islands and land reclamation projects will affect the accuracy of coastline extraction. Summary of the invention

[0004] In view of this, the present application provides a real-time coastline extraction method, which solves the problems in the prior art and realizes real-time extraction of coastlines in complex land and sea scenes.

[0005] The real-time coastline extraction method provided in this application adopts the following technical solution:

[0006] A real-time coastline extraction method comprises the following steps:

[0007] Step 1: Marking based on the existing coastline;

[0008] Step 2, generating training and test data, the training data includes a labeled coastline map and an unlabeled coastline map, and the test data includes a labeled coastline map and an unlabeled coastline map;

[0009] Step 3: Construction and training of deep learning network based on semantic segmentation;

[0010] The step 3 comprises:

[0011] Network initialization parameter settings;

[0012] The network initialization module generates a feature map for the network input training data, fuses the feature maps generated by the pooling layer and the convolution layer, and outputs data N1;

[0013] The network downsampling module 1 segments the image of the training data and extracts the feature information of the image;

[0014] The network downsampling module 2 performs feature extraction on the image generated by the network downsampling module 1;

[0015] The network upsampling process restores the resolution of the image of the network downsampling module 2;

[0016] Train and generate a coastline extraction network;

[0017] Step 4: Use the coastline extraction network to perform real-time coastline extraction on the test data.

[0018] Optionally, step 1 includes: interactive annotation based on radar echo data and electronic map prior information.

[0019] Optionally, the radar echo is subjected to pulse compression and a constant false alarm detection method is used to obtain a radar echo coastline, and the radar echo coastline and map prior information are combined using a four-neighborhood labeling method to obtain an interactive labeling result.

[0020] Optionally, step 2 includes performing a radar N-row scan data echo data matrix DataA and a corresponding interactive annotation result LandBoundFixAvg i , i=1:N is randomly selected to obtain training data and test data in a ratio of 5:1; the training data includes the coastline labeled map and the coastline unlabeled map, and the test data includes the coastline labeled map and the coastline unlabeled map.

[0021] Optionally, the network downsampling module 1 includes a stage 1 and a stage 2;

[0022] The stage 1 includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size is stride1=2, stride2=1, stride2=1, and the number of convolution kernels in each convolution layer is 32; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 32;

[0023] Input N1 to the two branches of the stage 1 respectively, perform sum operation on the outputs of the two branches and output N2;

[0024] Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations;

[0025] The second stage includes two branches, the two branches input N2 respectively, one branch directly outputs N2, and the other branch includes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64;

[0026] The output of the two branches is added and the output is N3, which has a size of (S1 / 4)×(S2 / 4)×N F1 ;

[0027] Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0028] Optionally, the network downsampling module 2 includes stage three and stage four;

[0029] The stage three includes a first operation, a second operation, a third operation, a fourth operation, a fifth operation, a sixth operation and a seventh operation;

[0030] The first operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the stride lengths are stride1=2, stride2=1, stride2=1, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the stride length is stride2=1, and the number of convolution kernels is 64, N3 is input to the two branches of the first operation respectively, the outputs of the two branches are added and output N4, wherein the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations;

[0031] The second operation includes two branches, the two branches input N4 respectively, one branch directly outputs N4, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N5. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations.

[0032] The third operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the first convolution kernel step size is stride1=2, the third convolution kernel step size is stride2=1, the second convolution kernel is a dilation rate dilation1 hollow convolution feature extraction, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N5 is input to the two branches of the third operation respectively, and the outputs of the two branches are summed to output N6, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations;

[0033] The fourth operation includes two branches, the two branches input N6 respectively, one branch directly outputs N6, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N7. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations.

[0034] The fifth operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the first convolution kernel step size is stride1=2, the third convolution kernel step size is stride2=1, the second convolution kernel is a hole convolution kernel with a dilation rate of dilation2 to extract features, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N7 is input to the two branches of the five operations respectively, and the outputs of the two branches are summed and output N8, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations;

[0035] The sixth operation includes two branches, the two branches input N8 respectively, one branch directly outputs N8, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N9. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations.

[0036] The seventh operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size of the first convolution kernel is stride1=2, the step size of the third convolution kernel is stride2=1, the second convolution kernel is a hole convolution kernel with a dilation rate of dilation3 to extract features, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3, the step size is stride2, and the number of convolution kernels is 64, N9 is input to the two branches of the five operations respectively, and the outputs of the two branches are added and output N10, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations;

[0037] The steps of stage 4 are to repeat stage 3, with a normalization layer BatchNorm and activation function ReLu between two adjacent convolution operations. The output of stage 4 is N11, with a size of (S1 / 8)×(S2 / 8)×N F2 .

[0038] Optionally, the network upsampling process includes transposed convolution, stage five, and the network full convolution layer FullConv;

[0039] The transposed convolution includes an input N11, a kernel size of Sizef1, a stride of stride2, a zero padding operation, and 32 convolution kernels. The transposed convolution output is used to obtain N12, and its size is: (S1 / 8)×(S2 / 8)×N F1 ;

[0040] Stage five includes two branches, which input N12 respectively. One branch directly outputs N12, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1. The number of convolution kernels is 16. The outputs of the two branches are added and output N13, with a size of (S1 / 8)×(S2 / 8)×N. F , where the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations;

[0041] The network full convolution layer FullConv includes a fully connected layer, followed by a softmax layer to achieve binary classification, and the output is a category number C=2, and the output size is S1×S2×C.

[0042] Optionally, the network initialization module in step 3 includes:

[0043] Input data matrix Traindata p ,p=1:P, the dimension is batchsize×S1×S2, and the network initialization module is output after passing through two branches;

[0044] A branch convolution layer, the convolution kernel size is Sizef1=3×3, the step size is stride1=2, the number of convolution kernels is 12, and then passes through the normalization layer BatchNorm and the activation function ReLu;

[0045] The other branch is the maximum pooling layer MaxP1, with a pooling size of SizeP1, a stride of stride1=2, and 4 convolution kernels;

[0046] The left and right channels of the output are merged to obtain N1, and the matrix dimension is: (S1 / 2)×(S2 / 2)×N F Among them, the number of output feature maps N F .

[0047] In summary, this application includes the following beneficial technical effects:

[0048] 1. By using radar echo data to observe complex environmental information, we have made breakthroughs in the interactive fusion annotation technology of real-time observation and map prior information, improved the real-time performance of data processing, and made breakthroughs in the construction of deep neural networks based on semantic segmentation. We have performed high-dimensional mapping of the boundary morphological features of coastline radar echo data, and improved the ability to extract coastlines in real time.

[0049] 2. This application improves the accuracy of coastline extraction for complex land and sea scenes in areas without map priors, and further enhances the intelligent perception capability of radar environment;

[0050] 3. The present invention uses radar measured data to achieve real-time extraction of coastlines in complex land and sea scenes, and can provide scene analysis capabilities for military areas without maps / prior information. It can also provide a basis for continuous observation and threat assessment of targets in sensitive military areas at the junction of land and sea. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 A flowchart of the real-time coastline extraction method for this application;

[0053] Figure 2 Extract coastline results for CFAR and electronic maps;

[0054] Figure 3 To adopt the coastline extraction results extracted by this application;

[0055] Figure 4 This is the network initialization module of this application;

[0056] Figure 5 It is the network structure component of the downsampling module of this application;

[0057] Figure 6 It is the network structure component of the basic feature module of this application;

[0058] Figure 7 To compare the coastline extraction results obtained in this application with those using the map prior method.

[0059] Description of reference numerals: DETAILED DESCRIPTION

[0060] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0061] The following describes the implementation methods of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work belong to the scope of protection of the present application.

[0062] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0063] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The drawings only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0064] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0065] An embodiment of the present application provides a real-time coastline extraction method.

[0066] like Figure 1 As shown, a real-time coastline extraction method comprises the following steps:

[0067] Step 1: Label based on the existing coastline.

[0068] Specifically, since the electronic map information currently used on aircraft has a certain update cycle and is subject to timeliness constraints, the electronic map and real-time radar echo data can be used for interactive annotation as the ground truth of the radar echo coastline data.

[0069] Interactive annotation based on radar echo data and electronic map prior information; the radar echo is subjected to pulse compression and constant false alarm detection method to obtain the radar echo coastline, which is then fused with the radar echo coastline and map prior information using the four-neighborhood labeling method to obtain the interactive annotation result.

[0070] Specifically, after the radar echo pulse compression, the search radar scan result of the i-th row is N in size. r ×M360×480 echo data matrix data i , radar N = 600 rows of scan data echo data matrix DataA is data i ,i=1:N data set: DataA={data1,data2...,data i ,....data N}, where N r The constant false alarm detection (CFAR) method is used to detect the data. i , i=1:N The coastline result obtained after processing is LandBFix i ,i=1:N,LandBFix i The matrix size is N r ×M. Figure 2 The solid line result is the coastline extraction result LandBFix1 obtained after the first row of scan echo data data1 is processed by CFAR.

[0071] The coastline extraction result LandBOrg obtained on DataA using map prior information i ,i=1:N,LandBOrg i The matrix size is N r ×M. Figure 2 The dotted line result is the coastline extraction result LandBOrg1 after the first row of scan echo data data1 is aligned with the map prior information.

[0072] LandBFix i ,i=1:N and LandBOrg i , i=1:N data is interactively annotated using the four-neighborhood labeling method to obtain the interactive annotation result LandBoundFixAvg i ,i=1:N,LandBoundFixAvg iThe matrix size is N r ×M.

[0073] Step 2: Generate training and test data. The training data includes the coastline labeled map and the coastline unlabeled map. The test data includes the coastline labeled map and the coastline unlabeled map. The radar N-row scanning data echo data matrix DataA and the corresponding interactive annotation result LandBoundFixAvg i , i=1:N is randomly selected to obtain training data and test data in a ratio of 5:1; the training data includes the coastline labeled map and the coastline unlabeled map, and the test data includes the coastline labeled map and the coastline unlabeled map

[0074] Step 3: Construction and training of deep learning network based on semantic segmentation.

[0075] Specifically, the step 3 includes:

[0076] Step 31, network initialization parameter setting; the training data is TrainD, and its matrix dimension is: N S ×S1×S2=500×360×480,N S =500 is the number of training samples, and the echo data matrix input into the network for each training is Traindata p ,p=1:P, P is the total number of training batches, P=300, batchsize=10 is the number of samples in each training batch, Traindata p The matrix dimension is S1×S2=360×480, where TrainD={Traindata1,Traindata2,...,Traindata p ,...,Traindata P}.

[0077] Step 32, the network initialization module generates a feature map for the network input training data, fuses the feature maps generated by the pooling layer and the convolution layer, and outputs data N1; specifically, Figure 4 As shown, the input data matrix Traindata p ,p=1:P, the dimension is batchsize×S1×S2=10×360×480, and the network initialization module is output after passing through two branches; Figure 4As shown in the figure, one branch convolution layer has a convolution kernel size of Sizef1=3×3, a stride of stride1=2, and 12 convolution kernels, and then passes through the normalization layer BatchNorm and the activation function ReLu; the other branch is the maximum pooling layer MaxP1, with a pooling size of SizeP1, a stride of stride1=2, and 4 convolution kernels; the left and right output channels are merged to obtain N1, and the matrix dimension is: (S1 / 2)×(S2 / 2)×N F Among them, the number of output feature maps N F , the matrix dimensions are: (S1 / 2)×(S2 / 2)×N F

[0078] =180×240×16, where the number of output feature maps is N F =16.

[0079] Step 33, the network downsampling module 1 segments the image of the training data and extracts the feature information of the image.

[0080] Specifically, the network downsampling module 1 includes stage 1 and stage 2:

[0081] like Figure 5 As shown, the stage one includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size is stride1=2, stride2=1, stride2=1, and the number of convolution kernels in each convolution layer is 32; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, the number of convolution kernels is 32, the maximum pooling layer is used to extract context information, and the convolution layer is used to realize channel conversion, so that the number of channels of the two branches is the same; N1 is input to the two branches of the stage one respectively, and the outputs of the two branches are added and output N2; wherein, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations.

[0082] like Figure 6As shown, the stage 2 includes two branches, the two branches input N2 respectively, one branch directly outputs N2, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64; the outputs of the two branches are added and output N3, with a size of (S1 / 4)×(S2 / 4)×N F1 , (S1 / 4)×(S2 / 4)×N F1 =90×120×32; among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0083] Step 34, the network downsampling module 2 performs feature extraction on the image generated by the network downsampling module 1.

[0084] Specifically, the network downsampling module 2 includes stage three and stage four:

[0085] The stage three includes a first operation, a second operation, a third operation, a fourth operation, a fifth operation, a sixth operation and a seventh operation.

[0086] like Figure 5 As shown, the first operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the stride lengths are stride1=2, stride2=1, stride2=1, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the stride length is stride2=1, and the number of convolution kernels is 64, N3 is input to the two branches of the first operation respectively, and the outputs of the two branches are added to output N4, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0087] like Figure 6As shown, the second operation includes two branches, the two branches input N4 respectively, one branch directly outputs N4, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N5. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0088] like Figure 5 As shown, the third operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the first convolution kernel step size is stride1=2, the third convolution kernel step size is stride2=1, the second convolution kernel is a dilation rate dilation1 hollow convolution feature extraction, dilation1=2; the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N5 is input to the two branches of the third operation respectively, the outputs of the two branches are added and output N6, wherein the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations.

[0089] like Figure 6 As shown, the fourth operation includes two branches, the two branches input N6 respectively, one branch directly outputs N6, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N7. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0090] like Figure 5As shown, the fifth operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the first convolution kernel step size is stride1=2, the third convolution kernel step size is stride2=1, the second convolution kernel is a hole convolution kernel with a dilation rate of dilation2 to extract features, dilation2=4, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N7 is input to the two branches of the five operations respectively, and the outputs of the two branches are added and output N8, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0091] like Figure 6 As shown, the sixth operation includes two branches, the two branches input N8 respectively, one branch directly outputs N8, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N9, wherein the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0092] like Figure 5 As shown, the seventh operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size of the first convolution kernel is stride1=2, the step size of the third convolution kernel is stride2=1, the second convolution kernel is a hole convolution kernel with a dilation rate of dilation3 to extract features, dilation3=8, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3, the step size is stride2, and the number of convolution kernels is 64, N9 is input to the two branches of the five operations respectively, the outputs of the two branches are added and output N10, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0093] The steps of stage 4 are to repeat stage 3, with a normalization layer BatchNorm and activation function ReLu between two adjacent convolution operations. The output of stage 4 is N11, with a size of (S1 / 8)×(S2 / 8)×N F2 , (S1 / 8)×(S2 / 8)×N F2 =45×60×64.

[0094] Step 35, the network upsampling process, restores the resolution of the image of the network downsampling module 2; and improves the network operation efficiency.

[0095] Specifically, the sampling process on the network includes transposed convolution, stage five, and the network full convolution layer FullConv:

[0096] The transposed convolution includes an input N11, a kernel size of Sizef1, a stride of stride2, a zero padding operation, and 32 convolution kernels. The transposed convolution output is used to obtain N12, and its size is: (S1 / 8)×(S2 / 8)×N F1 , (S1 / 8)×(S2 / 8)×N F1 =90×120×32.

[0097] Stage five includes two branches, which input N12 respectively. One branch directly outputs N12, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1. The number of convolution kernels is 16. The outputs of the two branches are added and output N13, with a size of (S1 / 8)×(S2 / 8)×N. F , (S1 / 8)×(S2 / 8)×N F =90×120×16, where a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations.

[0098] The network full convolution layer FullConv includes a fully connected layer, followed by a softmax layer to achieve binary classification, and the output is a category number C=2, and the output size is S1×S2×C, S1×S2×C=360×480×2.

[0099] Step 36, train and generate the coastline extraction network; use the Adam optimizer to train the network, select the cross entropy as the loss function, the training data TrainD, the label TrainL, and the initial learning rate lr=0.01 to obtain the corresponding network parameters and the trained model Model.

[0100] Step 4: Use the coastline extraction network to extract the coastline of the test data in real time. Use the trained model Model to extract the coastline of the test data TestD in real time. Figure 3 The coastline extraction result obtained by the method of the present invention is: Figure 7 The coastline extraction results obtained by the method of the present invention are compared with those obtained by using a map priori method, which illustrates the effectiveness of the proposed method.

[0101] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A real-time coastline extraction method, characterized in that: The steps include: Step 1, based on the existing coastline annotation: Based on the interactive annotation of radar echo data and electronic map prior information, the radar echo is subjected to pulse compression and the constant false alarm detection method is used to obtain the radar echo coastline, and the radar echo coastline and map prior information are combined using the four-neighborhood marking method to obtain the interactive annotation result; Step 2, generating training and test data, the training data includes a labeled coastline map and an unlabeled coastline map, and the test data includes a labeled coastline map and an unlabeled coastline map; Step 3: Construction and training of deep learning network based on semantic segmentation; The step 3 comprises: Network initialization parameter settings; The network initialization module generates a feature map for the network input training data, fuses the feature maps generated by the pooling layer and the convolution layer, and outputs data N1; The network downsampling module 1 segments the image of the training data and extracts the feature information of the image; The network downsampling module 2 performs feature extraction on the image generated by the network downsampling module 1; The network upsampling process restores the resolution of the image of the network downsampling module 2; Train and generate a coastline extraction network; Step 4, using the coastline extraction network to extract the coastline of the test data in real time; The network downsampling module 1 includes stage 1 and stage 2; The stage 1 includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size is stride1=2, stride2=1, stride2=1, and the number of convolution kernels in each convolution layer is 32; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 32; Input N1 to the two branches of the stage 1 respectively, perform sum operation on the outputs of the two branches and output N2; Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations; The second stage includes two branches, the two branches input N2 respectively, one branch directly outputs N2, and the other branch includes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64; The output of the two branches is added and the output is N3, which has a size of (S1 / 4)×(S2 / 4)×N F1 ; Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations; The network downsampling module 2 includes stage three and stage four; The stage three includes a first operation, a second operation, a third operation, a fourth operation, a fifth operation, a sixth operation and a seventh operation; The first operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel size of each convolution operation is Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size is stride1=2, stride2=1, stride2=1, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N3 is input to the two branches of the first operation respectively, the outputs of the two branches are added and output N4, wherein the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations; The second operation includes two branches, the two branches input N4 respectively, one branch directly outputs N4, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N5. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations. The third operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the first convolution kernel step size is stride1=2, the third convolution kernel step size is stride2=1, the second convolution kernel is a dilation rate dilation1 hollow convolution feature extraction, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N5 is input to the two branches of the third operation respectively, and the outputs of the two branches are summed to output N6, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations; The fourth operation includes two branches, the two branches input N6 respectively, one branch directly outputs N6, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N7. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations. The fifth operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the first convolution kernel step size is stride1=2, the third convolution kernel step size is stride2=1, the second convolution kernel is a dilation rate of dilation2 hollow convolution kernel feature extraction, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3=1×1, the step size is stride2=1, and the number of convolution kernels is 64, N7 is input to the two branches of the five operations respectively, and the outputs of the two branches are summed and output N8, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations; The sixth operation includes two branches, the two branches input N8 respectively, one branch directly outputs N8, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1, and the number of convolution kernels is 64. The outputs of the two branches are added and output N9. Among them, the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for the convolution operations. The seventh operation includes two branches, one branch undergoes three convolution operations in sequence, the convolution kernel sizes of each convolution operation are Sizef2=3×3, Sizef1=3×3, Sizef3=1×1, the step size of the first convolution kernel is stride1=2, the step size of the third convolution kernel is stride2=1, the second convolution kernel is a hole convolution kernel with a dilation rate of dilation3 to extract features, and the number of convolution kernels in each convolution layer is 64; the other branch includes a maximum pooling layer MaxP1 and a convolution layer, the size of the convolution layer is Sizef3, the step size is stride2, and the number of convolution kernels is 64, N9 is input to the two branches of the five operations respectively, and the outputs of the two branches are added and output N10, wherein a normalization layer BatchNorm and an activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for all convolution operations; The steps of stage 4 are to repeat stage 3, with a normalization layer BatchNorm and activation function ReLu between two adjacent convolution operations. The output of stage 4 is N11, with a size of (S1 / 8)×(S2 / 8)×N F2 .

2. The real-time coastline extraction method according to claim 1, characterized in that: The step 2 includes: performing an interactive annotation on the radar N-row scanning data echo data matrix DataA and the corresponding interactive annotation result LandBoundFixAvg i , i=1:N is randomly selected to obtain training data and test data in a ratio of 5:1; the training data includes the coastline labeled map and the coastline unlabeled map, and the test data includes the coastline labeled map and the coastline unlabeled map.

3. The real-time coastline extraction method according to claim 1, characterized in that: The sampling process on the network includes transposed convolution, stage five and the network full convolution layer FullConv; The transposed convolution includes an input N11, a kernel size of Sizef1, a stride of stride2, a zero padding operation, and 32 convolution kernels. The transposed convolution output is used to obtain N12, and its size is: (S1 / 8)×(S2 / 8)×N F1 ; Stage five includes two branches, which input N12 respectively. One branch directly outputs N12, and the other branch includes three convolution operations in sequence. The convolution kernel size of each convolution operation is Sizef3=1×1, Sizef1=3×3, Sizef3=1×1, and the step size is stride2=1, stride2=1, stride2=1. The number of convolution kernels is 16. The outputs of the two branches are added and output N13, with a size of (S1 / 8)×(S2 / 8)×N. F , where the normalization layer BatchNorm and the activation function ReLu are performed between two adjacent convolution operations, and zero padding operations are used for convolution operations; The network full convolution layer FullConv includes a fully connected layer, followed by a softmax layer to achieve binary classification, and the output is a category number C=2, and the output size is S1×S2×C.

4. The real-time coastline extraction method according to claim 1, characterized in that: The network initialization module in step 3 includes: Input data matrix Traindata p ,p=1:P, the dimension is batchsize×S1×S2, and the network initialization module is output after passing through two branches; A branch convolution layer, the convolution kernel size is Sizef1=3×3, the step size is stride1=2, the number of convolution kernels is 12, and then passes through the normalization layer BatchNorm and the activation function ReLu; The other branch is the maximum pooling layer MaxP1, with a pooling size of SizeP1, a stride of stride1=2, and 4 convolution kernels; The left and right channels of the output are merged to obtain N1, and the matrix dimension is: (S1 / 2)×(S2 / 2)×N F Among them, the number of output feature maps N F .

Images