CN114187527A - Transfer learning ship target segmentation method based on linear heating and snapshot integration - Google Patents

Transfer learning ship target segmentation method based on linear heating and snapshot integration Download PDF

Info

Publication number
CN114187527A
CN114187527A CN202111427112.1A CN202111427112A CN114187527A CN 114187527 A CN114187527 A CN 114187527A CN 202111427112 A CN202111427112 A CN 202111427112A CN 114187527 A CN114187527 A CN 114187527A
Authority
CN
China
Prior art keywords
data set
model
segmentation
source domain
domain
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111427112.1A
Other languages
Chinese (zh)
Other versions
CN114187527B (en
Inventor
张修社
韩春雷
亓子龙
任子豪
孙晓龙
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
CETC 20 Research Institute
Original Assignee
CETC 20 Research Institute
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by CETC 20 Research Institute filed Critical CETC 20 Research Institute
Priority to CN202111427112.1A priority Critical patent/CN114187527B/en
Publication of CN114187527A publication Critical patent/CN114187527A/en
Application granted granted Critical
Publication of CN114187527B publication Critical patent/CN114187527B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention provides a migration learning ship target segmentation method based on linear heating and snapshot integration, which comprises the steps of taking a data set with uniform size distribution of ship targets as a source domain data set, taking a data set with verified segmentation effect as a target domain data set, constructing a global attention-based coding and decoding network of a source domain, obtaining a model with the highest segmentation accuracy on a source domain test set, constructing a global attention-based coding and decoding network of a target domain, obtaining a final segmentation model, and testing the target domain test data set by using the final segmentation model to obtain a final segmentation result. The method can realize the ship target segmentation under fewer labels on the target domain, solves the problem of large demand of the labels of the target domain to a certain extent, integrates the optimal model in each cycle of the cyclic cosine annealing, avoids the condition that the training is unstable when the data volume of the target domain is less so as to cause negative migration, and enhances the robustness of the model.

Description

Transfer learning ship target segmentation method based on linear heating and snapshot integration
Technical Field
The invention relates to the field of image segmentation, in particular to a transfer learning method which can be used for intermediate processing of SAR image interpretation.
Background
In recent years, with the development of synthetic aperture radar systems, acquired information is gradually transferred from land to the sea, and how to solve the problem of small sample ship target segmentation of the SAR image becomes an urgent need to be solved at present. In recent years, with the excellent performance of deep learning in the fields of computer vision, speech signal processing, natural language processing and the like, how to combine the deep learning method with the SAR image ship target segmentation problem also becomes a hotspot problem in the SAR image processing field nowadays. The deep learning method is characterized in that inherent attribute characteristics of training data are continuously mined through a thought of training and learning layer by layer, and further, abstract representation of the data is realized.
The Chenyangtang et al, in the article "remote sensing image sea surface ship detection research based on depth semantic segmentation", proposes a segmentation method, which is based on the ResNet architecture, firstly, the remote sensing image is taken as input through a depth convolution neural network, the image is roughly segmented, then through an improved full-connection conditional random field, a conditional random field is established by using Gauss paired potential and average field approximate theorem as an output through a recurrent neural network, thereby realizing the end-to-end connection.
Wan 281569, in a paper, "Multi-Scale CNN method in image segmentation", proposes an SAR image ship detection segmentation method based on a three-dimensional void convolutional neural network, and the method constructs a three-dimensional image block based on multi-scale by adding image wavelet features, and improves the capability of extracting target global features and local features by using the three-dimensional image block as the input of the three-dimensional void convolutional neural network. The three-dimensional cavity convolution neural network adopts an end-to-end network structure, the network output is the final output result, and the model is convenient to use and has higher efficiency.
However, the method is limited by the problem that the SAR image data scale is small in a complex large scene, the model often has the condition of insufficient generalization capability, namely the model often has superior performance on source domain data, but has the phenomenon of performance decline on a target domain, and a means for formally solving the problem through transfer learning is provided, perhaps a universal model which is universal and cannot be obtained, but a model with acceptable performance can be reformed and adapted to realize the personalized task of a specific scene. Therefore, the transfer learning is widely applied to solving the problem of detection and identification of the heterogeneous images.
However, since the parameters such as transfer learning and learning rate have close relationship, the method for directly transferring the image is not well suitable for SAR image ship target segmentation.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a migration learning ship target segmentation method based on linear heating and snapshot integration, which is a migration learning ship target segmentation method based on linear heating, cyclic cosine annealing and snapshot integration, and improves the segmentation effect on the premise of reducing the number of labeled data required by data on a target domain.
The technical scheme adopted by the invention for solving the technical problem comprises the following steps:
(1) taking a data set with uniform size distribution of ship targets as a source domain data set, and dividing the source domain data set into a source domain training data set and a source domain testing data set in proportion;
(2) taking a data set with a verification segmentation effect as a target domain data set, and dividing the target domain data set into a target domain training data set, a target domain verification data set and a target domain test data set in proportion;
(3) constructing a global attention coding and decoding network of a source domain, training the global attention coding and decoding network by using a source domain training data set, keeping a model with the highest segmentation accuracy on a source domain test set after each iteration, and obtaining the model with the highest segmentation accuracy on the source domain test set after the maximum iteration times are reached; the method comprises the following specific steps:
(3a) constructing a global attention-based coding and decoding network of a source domain, wherein the global attention-based coding and decoding network comprises a cascaded input layer, a cascaded coding layer, a cascaded decoding layer and a cascaded output layer, and the coding layer and the decoding layer are connected by a global attention module;
(3b) setting a training optimizer as SGD, and adopting cross entropy loss as a loss function;
(3c) training a global attention-based coding and decoding network of a source domain by using a source domain training data set by adopting a small-batch gradient descent algorithm, testing by using a source domain test data set after each iteration in the training process until the maximum iteration number is reached, and obtaining an optimal model M on the model most source domain with the best test result on the source domain test data setS；
(4) The global attention coding and decoding network of the target domain is constructed in the same way as the global attention coding and decoding network of the source domain, and the global attention coding and decoding network model M of the target domainTBy the parameter MSInitializing the parameters;
(5) the method comprises the steps of setting a learning rate in an early stage of training by using a linear heating strategy in a learning rate setting method, setting a learning rate in a later stage of training by using a cyclic cosine annealing strategy in the learning rate setting method, integrating by using a snapshot integration strategy in an integration method, wherein a model reserved in the snapshot integration strategy is a model with the highest segmentation accuracy on a target domain verification set in each cycle in the cyclic cosine annealing strategy, an optimizer adopts SGD (generalized maximum likelihood), a loss function adopts a cross entropy loss function, and the last three models reserved in the snapshot integration strategy are obtained final segmentation models ME；
(6) The most usedFinal segmentation model MEAnd testing the target domain test data set to obtain a final segmentation result.
The source domain data set is divided into a source domain training data set and a source domain testing data set according to the proportion of 4:1, and the target domain data set is divided into a target domain training data set, a target domain verification data set and a target domain testing data set according to the proportion of 4:1: 1;
the expression of the cross entropy loss function is as;
Figure BDA0003379012970000031
wherein X represents the number of samples, C represents the number of classes,
Figure BDA0003379012970000032
labels indicating sample x, when sample x is of class c
Figure BDA0003379012970000033
Is 1, is other
Figure BDA0003379012970000034
Is a non-volatile organic compound (I) with a value of 0,
Figure BDA0003379012970000035
representing the probability that sample x is predicted as class c.
The optimal model MSComprises the following steps: keeping the model with the highest segmentation accuracy on the source domain test set after each iteration until the maximum iteration number E is reachedSThen, obtaining a model with the highest segmentation accuracy on the source domain test data set, and taking the model as an optimal model M on the source domainS；
The specific steps of the step (5) are as follows:
(5a) setting the number of iteration rounds E of the initial linear heatinglInitial learning rate lrlInitial learning rate lr of cyclic cosine annealing in snapshot integration strategymTotal number of rounds of iteration E during transformation periodmAnd the number of cycles used, n;
(5b) fine tuning training using k% of the target domain training dataset, front ElSetting the learning rate using a linear heating strategy in round iterations, followed by EmIn the round iteration, a cyclic cosine annealing strategy is used for setting the learning rate, and in the cyclic cosine annealing stage, every EmSaving current E after n iterationsmModel with highest segmentation accuracy on target domain verification set in n iterations
Figure BDA0003379012970000036
(5c) Selection using snapshot integration strategy
Figure BDA0003379012970000041
The three models are used as the final segmentation model MEWherein i is n-3, n-2, n-1, and when in segmentation, the segmentation results of the three models are averaged to obtain a final segmentation model ME。
The values in the step (5a) are as follows: number of iteration rounds E of initial linear heatinglAn initial learning rate lr of 10l0.0001, initial learning rate lr of cyclic cosine annealing in snapshot integration strategym0.01, total number of rounds of iteration E during the transformation periodmThe number of cycles n used was 5, which was 50.
The invention has the beneficial effects that:
1) the heterogeneous SAR image ship target under a small number of labels can be segmented.
The invention provides a migration learning ship target segmentation method based on linear heating and snapshot integration, which is characterized in that a data set relatively rich in ship targets is used as source domain data, the data set is used for training, and the obtained training model can realize ship target segmentation under fewer labels on a target domain under the fine adjustment of a small amount of data of the target domain, so that the problem of large demand of the labels of the target domain is solved to a certain extent.
2) Using multiple strategies enhances the robustness of the model.
In the invention, for the fine tuning part after the migration, the strategy of linear heating and cyclic cosine annealing is used for the learning rate of the part, and the snapshot integration strategy is used, so that the optimal model in each cycle of the cyclic cosine annealing is integrated, the condition that the training is unstable when the data volume of the target domain is less so as to cause negative migration is avoided, and the robustness of the model is enhanced.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a ship size distribution condition in an SAR image of a Qingdao area in the invention;
FIG. 3 is a graph of the effect of different migration data volumes when the hong Kong area is used as the target domain in the present invention; where, graph (a) is the 0% data migration using 0% data migration, graph (b) is the 0% data migration using 2% data migration, graph (c) is the 0% data migration using 3% data migration, graph (d) is the 0% data migration using 4% data migration, graph (e) is the 0% data migration using 5% data migration, and graph (f) is the 0% data migration using 10% data migration.
FIG. 4 is a graph illustrating the effect of different migration data volumes when the Shanghai region is used as the target domain; where, graph (a) is the 0% data migration using 0% data migration, graph (b) is the 0% data migration using 2% data migration, graph (c) is the 0% data migration using 3% data migration, graph (d) is the 0% data migration using 4% data migration, graph (e) is the 0% data migration using 5% data migration, and graph (f) is the 0% data migration using 10% data migration.
FIG. 5 is a graph showing the effect of different migration data amounts when IsteBoolean harbor is used as the target domain in the present invention. Where, graph (a) is the 0% data migration using 0% data migration, graph (b) is the 0% data migration using 2% data migration, graph (c) is the 0% data migration using 3% data migration, graph (d) is the 0% data migration using 4% data migration, graph (e) is the 0% data migration using 5% data migration, and graph (f) is the 0% data migration using 10% data migration.
Detailed Description
The invention is further illustrated with reference to the following figures and examples.
Referring to fig. 1, the implementation steps of the present invention include the following:
step 1, taking a data set rich in ship targets as a source domain data set, and dividing the data set into a source domain training data set and a source domain testing data set according to a ratio of 4: 1.
Step 2, taking the data set with the verified segmentation effect as a target domain data set, and dividing the data set into a target domain training data set, a target domain verification data set and a target domain test data set according to the ratio of 4:1: 1;
and 3, constructing a global attention coding and decoding network of the source domain, and training by using a source domain training data set to obtain an optimal model on the source domain verification set.
(3.1) constructing a global attention coding and decoding network of a source domain, wherein the global attention coding and decoding network comprises an input layer, an encoding layer, a decoding layer, an output layer and a global attention module;
(3.1a) in an input layer, performing wavelet decomposition on an input image to obtain a plurality of images with the same size as the input image, and overlapping the input image and each image subjected to wavelet decomposition to construct a 3D input image block;
(3.1b) in the coding layer, extracting features from the image blocks obtained by the input layer to obtain coding features of the input image;
(3.1c) the global attention module fuses the high-dimensional semantic features and the low-dimensional position features to obtain features fusing semantic and position information;
(3.1d) in a decoding layer, decoding the features passing through the attention module to obtain a feature map with continuously increased scale;
(3.1e) obtaining a pixel-level segmentation result in the output layer through the convolution layer and the softmax layer;
(3.1f) the network consists of a cascade of input, encoding, global attention, decoding and output layers.
(3.2) randomly initializing network parameters, setting a training optimizer as SGD, wherein a cross entropy loss function is adopted as a loss function, and the cross entropy loss function is expressed as;
Figure BDA0003379012970000061
wherein X represents the number of samples, C represents the number of classes,
Figure BDA0003379012970000062
labels indicating sample x, when sample x is of class c
Figure BDA0003379012970000063
Is 1, is other
Figure BDA0003379012970000064
Is a non-volatile organic compound (I) with a value of 0,
Figure BDA0003379012970000065
represents the probability that sample x is predicted as class c;
(3.3) training the network, testing by using the source domain test data set after each iteration in the training process, comparing with a historical result, and keeping a model with the best test result;
(3.4) repeating (3.3) until the iteration reaches the maximum iteration number, terminating training and obtaining the optimal model M on the verification setS。
Step 4, constructing a global attention coding and decoding network of the target domain, and a model M thereofTBy the parameter MSIs initialized.
Step 5, migration learning is carried out by using a linear heating and snapshot integration strategy, an SGD (generalized regression) is adopted by an optimizer, a cross entropy loss function is adopted by the loss function, a plurality of models of snapshot integration are obtained by a cyclic cosine annealing strategy, and a final segmentation model M is obtainedE。
(5.1) setting the number of iteration rounds E of the initial linear heatinglInitial learning rate lrlAnd initial learning rate lr of cyclic cosine annealingmTotal number of rounds of iteration E during transformation periodmAnd useThe number of cycles n;
(5.2) setting the learning rate to change according to the settings of linear heating and cyclic cosine annealing;
(5.2a) in the early stage of training, the learning rate is changed in a linear heating mode, and the learning rate is slowly increased to lr from zerolThe mathematical expression of the learning rate is:
Figure BDA0003379012970000066
wherein, lrlIs the maximum learning rate of the linear heating stage, s is the current iteration step number, ElThe total number of iteration rounds of the linear heating stage is shown, and S is the number of iteration steps of each iteration round.
(5.2b) in the later stage of training, the change of the learning rate meets a cosine period attenuation function, the learning rate is suddenly increased from the minimum value to the initially set maximum learning rate after each generation of training, and the mathematical expression of the learning rate is as follows:
Figure BDA0003379012970000067
wherein, lrmFor the initially set maximum learning rate, t is the current iteration number, t is 0 when the linear heating stage is finished, EmIs the total number of iterations during the decay of the cosine period, and n is the number of periods.
(5.3) performing fine tuning training by using k% of a target domain training data set, wherein the learning rate is changed according to the settings of linear heating and cyclic cosine annealing, and after the linear heating stage, E is passed every timemN iterations, save current EmModel for optimal representation on verification set in n iterations
Figure BDA0003379012970000071
(5.3) selection Using Snapshot Integrated policy
Figure BDA0003379012970000072
The three models are used for integration to obtain a final segmentation model MEWherein the segmentation model M is integrated from snapshotsECan be represented as;
Figure BDA0003379012970000073
wherein M isERepresenting the final integrated model and x representing the input image to be segmented.
Step 6, using the model MEAnd testing the target domain test data set to obtain a final segmentation result.
(6.1) sequentially taking a sample from the target domain test set and inputting the sample into the trained integrated model to obtain a segmentation result pred corresponding to the sample;
and (6.2) repeating (6.1) until all the query images of the target domain test set obtain the segmentation result, and ending the test.
The effects of the present invention can be further illustrated by the following simulations.
Simulation data
The general data set comprises a Qingdao area data set, a hong Kong area data set, a Shanghai area data set and an Isteboolean harbor data set, the ship target size distribution of the Qingdao area is relatively average, large and small ship targets are few, the distribution graph of the Qingdao area ship target size is shown in figure 2, the size distribution greatly helps the training of the model, and therefore the data set of the Qingdao area is selected as a source domain data set, and the data sets of other three areas are respectively selected as target domain data sets.
Emulated content
The method adopts three segmentation methods based on a global attention coding and decoding network (GAM-EDNet) and the existing PSPNet and DCNN as comparison methods, carries out comparison experiments with four methods based on the global attention coding and decoding network under the migration learning in the invention, has the same training data in the experiments, respectively carries out verification experiments by using data sets of hong Kong area, Shanghai area and Instantbuer area, respectively verifies the effects of the method in the invention under the condition that the migration data volume is 0%, 2%, 3%, 4%, 5% and 10%, respectively uses the cross-over ratio, the weighted cross-over ratio and the Kappa coefficient as evaluation indexes, and shows a segmentation effect graph as shown in FIG. 3, wherein FIG. 3 is an effect graph of different migration data volumes when the hong Kong area is taken as a target domain, FIG. 4 is an effect graph of different migration data volumes when the Shanghai area is taken as a target domain, fig. 5 is a graph showing the effect of different migration data amounts of the method when istembushbound is used as a target domain, and tables 1 and 2 are a comparison of results of different algorithms in the case that the target domain is hong kong and results of the method under different migration data amounts, respectively, wherein 10% of target domain data are used for training in table 1.
TABLE 1 comparison of different algorithm results in hong Kong area
Figure BDA0003379012970000081
TABLE 2 results of the method in hong Kong area under different migration data volume
Figure BDA0003379012970000082
Tables 3 and 4 are the comparison of the results of different algorithms and the results of the method at different migration data volumes when the target domain is the Shanghai region, respectively, wherein 10% of the target domain data is used for training in table 3.
TABLE 3 comparison of different algorithm results in Shanghai region
Figure BDA0003379012970000083
TABLE 4 results of the method in hong Kong area under different migration data volume
Figure BDA0003379012970000091
Tables 5 and 6 are a comparison of the results of the different algorithms in the case of the target domain being the instam harbor region and the results of the method at different migration data volumes, respectively, wherein 10% of the target domain data was used for training in table 5.
TABLE 5 comparison of different algorithm results for Isteboolean harbor region
Figure BDA0003379012970000092
TABLE 4 results of the method in IstanBoolean harbor at different migration data volumes
Figure BDA0003379012970000093
Simulation effect analysis
From tables 1, 3 and 5, it can be seen that the method can achieve the optimal effect under the condition of using 10% of training data of the target domain compared with other methods, thereby proving the better effect of the method under the condition of less training data of the target domain.
As can be seen from tables 2, 4, and 6, in the three regions, the performance of segmentation becomes better as the migration amount of the target domain data increases, and as can also be seen from fig. 4, the region consistency of the segmentation result becomes better as the migration amount of the target domain data increases.
From the simulation results, the method can achieve a good effect under the condition that the target domain training data are less, and fully explains the effectiveness of the migration learning and the linear heating and snapshot integration strategy used in the method.

Claims (6)

1. A migration learning ship target segmentation method based on linear heating and snapshot integration is characterized by comprising the following steps:
(1) taking a data set with uniform size distribution of ship targets as a source domain data set, and dividing the source domain data set into a source domain training data set and a source domain testing data set in proportion;
(2) taking a data set with a verification segmentation effect as a target domain data set, and dividing the target domain data set into a target domain training data set, a target domain verification data set and a target domain test data set in proportion;
(3) constructing a global attention coding and decoding network of a source domain, training the global attention coding and decoding network by using a source domain training data set, keeping a model with the highest segmentation accuracy on a source domain test set after each iteration, and obtaining the model with the highest segmentation accuracy on the source domain test set after the maximum iteration times are reached; the method comprises the following specific steps:
(3a) constructing a global attention-based coding and decoding network of a source domain, wherein the global attention-based coding and decoding network comprises a cascaded input layer, a cascaded coding layer, a cascaded decoding layer and a cascaded output layer, and the coding layer and the decoding layer are connected by a global attention module;
(3b) setting a training optimizer as SGD, and adopting cross entropy loss as a loss function;
(3c) training a global attention-based coding and decoding network of a source domain by using a source domain training data set by adopting a small-batch gradient descent algorithm, testing by using a source domain test data set after each iteration in the training process until the maximum iteration number is reached, and obtaining an optimal model M on the model most source domain with the best test result on the source domain test data setS；
(4) The global attention coding and decoding network of the target domain is constructed in the same way as the global attention coding and decoding network of the source domain, and the global attention coding and decoding network model M of the target domainTBy the parameter MSInitializing the parameters;
(5) the learning rate of the training early stage is set by using a linear heating strategy in a learning rate setting method, the learning rate of the training later stage is set by using a cyclic cosine annealing strategy in the learning rate setting method, the integration is carried out by using a snapshot integration strategy in an integration method, a model reserved in the snapshot integration strategy is a model with the highest segmentation accuracy on a target domain verification set in each cycle in the cyclic cosine annealing strategy, an optimizer adopts SGD, a cross entropy loss function is adopted as the loss function, and the snapshot integration function is adopted as the modelThe last three models retained in the strategy are the final segmentation model ME；
(6) Using the final segmentation model MEAnd testing the target domain test data set to obtain a final segmentation result.
2. The migration learning ship target segmentation method based on linear heating and snapshot integration according to claim 1, characterized in that:
the source domain data set is divided into a source domain training data set and a source domain testing data set according to the proportion of 4:1, and the target domain data set is divided into a target domain training data set, a target domain verification data set and a target domain testing data set according to the proportion of 4:1: 1.
3. The migration learning ship target segmentation method based on linear heating and snapshot integration according to claim 1, characterized in that:
the expression of the cross entropy loss function is as;
Figure FDA0003379012960000021
wherein X represents the number of samples, C represents the number of classes,
Figure FDA0003379012960000022
labels indicating sample x, when sample x is of class c
Figure FDA0003379012960000023
Is 1, is other
Figure FDA0003379012960000024
Is a non-volatile organic compound (I) with a value of 0,
Figure FDA0003379012960000025
representing the probability that sample x is predicted as class c.
4. The migration learning ship target segmentation method based on linear heating and snapshot integration according to claim 1, characterized in that:
the optimal model MSComprises the following steps: keeping the model with the highest segmentation accuracy on the source domain test set after each iteration until the maximum iteration number E is reachedSThen, obtaining a model with the highest segmentation accuracy on the source domain test data set, and taking the model as an optimal model M on the source domainS。
5. The migration learning ship target segmentation method based on linear heating and snapshot integration according to claim 1, characterized in that:
the specific steps of the step (5) are as follows:
(5a) setting the number of iteration rounds E of the initial linear heatinglInitial learning rate lrlInitial learning rate lr of cyclic cosine annealing in snapshot integration strategymTotal number of rounds of iteration E during transformation periodmAnd the number of cycles used, n;
(5b) fine tuning training using k% of the target domain training dataset, front ElSetting the learning rate using a linear heating strategy in round iterations, followed by EmIn the round iteration, a cyclic cosine annealing strategy is used for setting the learning rate, and in the cyclic cosine annealing stage, every EmSaving current E after n iterationsmModel with highest segmentation accuracy on target domain verification set in n iterations
Figure FDA0003379012960000026
(5c) Selection using snapshot integration strategy
Figure FDA0003379012960000031
The three models are used as the final segmentation model MEWherein i is n-3, n-2, n-1, and when in segmentation, the segmentation results of the three models are averaged to obtain a final segmentation model ME。
6. The migration learning ship target segmentation method based on linear heating and snapshot integration according to claim 5, wherein:
the values in the step (5a) are as follows: number of iteration rounds E of initial linear heatinglAn initial learning rate lr of 10l0.0001, initial learning rate lr of cyclic cosine annealing in snapshot integration strategym0.01, total number of rounds of iteration E during the transformation periodmThe number of cycles n used was 5, which was 50.
CN202111427112.1A 2021-11-28 2021-11-28 Migration learning ship target segmentation method based on linear heating and snapshot integration Active CN114187527B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111427112.1A CN114187527B (en) 2021-11-28 2021-11-28 Migration learning ship target segmentation method based on linear heating and snapshot integration

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111427112.1A CN114187527B (en) 2021-11-28 2021-11-28 Migration learning ship target segmentation method based on linear heating and snapshot integration

Publications (2)

Publication Number Publication Date
CN114187527A true CN114187527A (en) 2022-03-15
CN114187527B CN114187527B (en) 2022-12-27

Family

ID=80602823

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111427112.1A Active CN114187527B (en) 2021-11-28 2021-11-28 Migration learning ship target segmentation method based on linear heating and snapshot integration

Country Status (1)

Country Link
CN (1) CN114187527B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359557A (en) * 2018-09-25 2019-02-19 东北大学 A kind of SAR remote sensing images Ship Detection based on transfer learning
CN112926547A (en) * 2021-04-13 2021-06-08 北京航空航天大学 Small sample transfer learning method for classifying and identifying aircraft electric signals
CN113610097A (en) * 2021-08-09 2021-11-05 西安电子科技大学 SAR ship target segmentation method based on multi-scale similarity guide network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359557A (en) * 2018-09-25 2019-02-19 东北大学 A kind of SAR remote sensing images Ship Detection based on transfer learning
CN112926547A (en) * 2021-04-13 2021-06-08 北京航空航天大学 Small sample transfer learning method for classifying and identifying aircraft electric signals
CN113610097A (en) * 2021-08-09 2021-11-05 西安电子科技大学 SAR ship target segmentation method based on multi-scale similarity guide network

Also Published As

Publication number Publication date
CN114187527B (en) 2022-12-27

Similar Documents

Publication Publication Date Title
Dai et al. Compressing neural networks using the variational information bottleneck
CN106203625B (en) A kind of deep-neural-network training method based on multiple pre-training
CN105138973A (en) Face authentication method and device
CN108985317A (en) A kind of image classification method based on separable convolution sum attention mechanism
CN109375186A (en) Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error
CN112541572A (en) Residual oil distribution prediction method based on convolutional encoder-decoder network
Zhou et al. Exploiting operation importance for differentiable neural architecture search
Meng et al. A Fast Recognition Algorithm of Online Social Network Images Based on Deep Learning.
CN109151727B (en) WLAN fingerprint positioning database construction method based on improved DBN
CN107273934A (en) A kind of figure clustering method merged based on attribute
CN112529068B (en) Multi-view image classification method, system, computer equipment and storage medium
CN111125411A (en) Large-scale image retrieval method for deep strong correlation hash learning
CN114491039B (en) Primitive learning few-sample text classification method based on gradient improvement
CN113222998A (en) Semi-supervised image semantic segmentation method and device based on self-supervised low-rank network
CN113420593A (en) Small sample SAR automatic target recognition method based on hybrid inference network
Lin et al. A frequency-domain convolutional neural network architecture based on the frequency-domain randomized offset rectified linear unit and frequency-domain chunk max pooling method
CN112464891A (en) Hyperspectral image classification method
de Bézenac et al. Optimal unsupervised domain translation
CN116258877A (en) Land utilization scene similarity change detection method, device, medium and equipment
CN116450941A (en) Book recommendation method and system based on Lorentz chart convolution network
CN116563683A (en) Remote sensing image scene classification method based on convolutional neural network and multi-layer perceptron
CN110188864B (en) Small sample learning method based on distribution representation and distribution measurement
CN114187527B (en) Migration learning ship target segmentation method based on linear heating and snapshot integration
Shu et al. Wasserstein distributional harvesting for highly dense 3D point clouds
CN116843544A (en) Method, system and equipment for super-resolution reconstruction by introducing hypersonic flow field into convolutional neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant