CN113361439B

CN113361439B - SAR image ship target identification method and system

Info

Publication number: CN113361439B
Application number: CN202110673322.2A
Authority: CN
Inventors: 徐从安; 李健伟; 姚力波; 王海洋; 吴俊峰; 孙炜玮; 苏航
Original assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Current assignee: School Of Aeronautical Combat Service Naval Aeronautical University Of Pla
Priority date: 2021-06-17
Filing date: 2021-06-17
Publication date: 2022-06-17
Anticipated expiration: 2041-06-17
Also published as: CN113361439A

Abstract

The invention relates to a method for identifying a ship target in an SAR image. The SAR image ship target identification method preprocesses training data through the proposed inter-class sample imbalance processing technology (including data enhancement based up-sampling processing and proportional generation batch method), improves the diversity of the training data, and simultaneously keeps the training data sent to a network balanced among classes; through the provided dense residual error network for SAR image ship target identification, the original characteristics can be reused while more new characteristics are learned; by the loss function based on the central loss, the parameters of the network model are adjusted, and the intra-class compactness and the inter-class separability are optimized simultaneously. Recognition results on OpenSARShip show that the designed dense residual network has higher accuracy, smaller model size and smaller calculation amount than the existing adopted neural network.

Description

SAR image ship target identification method and system

Technical Field

The invention relates to the technical field of target identification and radar remote sensing, in particular to a method and a system for identifying a ship target in an SAR image.

Background

Synthetic Aperture Radar (SAR) uses the movement of a remote sensing platform to install an antenna with a small Aperture at the side of the platform and forms an equivalent antenna with a large Aperture through movement, thereby achieving the purpose of improving the azimuth resolution. The all-weather sensor can generate high-resolution SAR images and is widely applied to the military and civil fields. With the increase of the available SAR images, the automatic and intelligent interpretation of the SAR images becomes more and more important, and ships as important military and civil targets are objects needing important attention, so that the SAR image ship target identification is researched in a large amount, and a lot of results are obtained. At present, the SAR image ship target identification method mostly refers to a general target identification method, and the development of the vein is from a traditional method to a CNN-based identification method. The traditional method comprises two parts of feature extraction and classifier design, when the features of the target in the SAR image are extracted, for small targets with unclear details, geometric features such as area, perimeter, length-width ratio and moment can be utilized, and for high-resolution targets with clear details, some higher-level features such as LBP and SIFT can be utilized.

In recent years, the CNN-based SAR image target identification method is researched more, wherein the research is more intensive in China's double-denier university, Shanghai's delivery, electronic institute of Chinese academy, remote sensing institute of Chinese academy, and naval aviation university, and the method has greater advantages compared with the traditional identification method. For example (Bentes C, Velotto D, Tings B. Ship Classification in Terrra AR-X Images With volumetric Neural networks IEEE Journal of scientific engineering.2017, PP (99):1-9.) four specially designed CNN models (named CNN-A, CNN-B, CNN-C and CNN-D, respectively) were proposed for the Classification of objects in SAR Images, With good results, but no data set was disclosed. In order to reduce the number of parameters, the paper (s.chen, h.wang, f.xu, and y-. q.jin.target classification using the deep correlation networks for SAR images [ J ]. IEEE trans.geosci.remote sens.2016, vol.54, No.8, pp.4806-4817.) proposes a-ConvNets, which only contains a few sparse connection layers, without full connection layers. However, the shallow CNNs used in the above are still far from the current more advanced CNNs such as ResNet and densneet in terms of feature expression. (Shao J, Qu C, Li J.A performance analysis of systematic neural network models in SAR target simulation [ C ].1-6.10.1109/BIGSARDATA.2017.8124917.) on the public data set MSATR, the identification performance of the classical CNN (including LeNet, AlexNet, ResNet and DesnseNet, etc.) is subjected to a comparative experiment, and the experimental result shows that the classification accuracy can easily reach more than 99%, which shows the strong performance of the SAR image target identification method. The currently disclosed data set for SAR image target identification is only available from OpenSARShip, but the data set has the problem of sample imbalance among categories, so that the identification performance is seriously reduced. And the existing classic CNN is specially designed for three-channel natural images and has the problems of model parameter redundancy and large calculated amount when being applied to SAR images.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a method and a system for identifying a ship target in an SAR image.

In order to achieve the purpose, the invention provides the following scheme:

a SAR image ship target identification method comprises the following steps:

adopting a batch generation method according to a proportion to construct a training sample data set;

acquiring an initial dense residual error network;

training the initial dense residual error network by adopting the training sample data set to obtain a trained dense residual error network;

acquiring SAR image data to be identified;

performing data expansion according to the SAR image data in an image enhancement mode to obtain expanded image data;

and inputting the expanded image data into the trained dense residual error network to obtain a ship target identification result.

Preferably, the constructing the training sample data set by using the batch generation according to the proportion method specifically includes:

acquiring the class of a ship;

selecting a set amount of sample data to form a batch based on each ship category;

and forming a training sample data set by different batches of sample data.

Preferably, the selecting a set number of sample data based on each ship category forms a batch, specifically including:

counting the number of samples of each ship category, and performing ascending arrangement according to the number of the samples;

selecting a set number of samples in each category to form a batch; and when the residual number of the samples in the current ship category is smaller than the set number, selecting the residual samples and the samples at the beginning of the sample sequence to form a batch, and repeating the steps until the specified iteration times are reached.

Preferably, the training the initial dense residual error network by using the training sample data set to obtain a trained dense residual error network specifically includes:

inputting the training sample data set into the initial dense residual error network for feature extraction, and outputting a predicted value of a category;

determining a loss value according to a predicted value and a true value in the training sample data set based on a loss function of an initial dense residual error network, updating model parameters of the initial dense residual error network by using a gradient descent method based on the loss value until convergence training is finished, and obtaining a trained dense residual error network; the loss function of the initial dense residual network is a loss function based on a central loss.

Preferably, the trained dense residual network comprises a convolutional layer, a pooling layer, a transformation layer, a classification layer and a plurality of dense residual modules; the dense residual error module is a feature learning module for aggregating the dense connection and the residual error connection.

According to the specific embodiment provided by the invention, the invention discloses the following technical effects:

according to the SAR image ship target identification method provided by the invention, training data are preprocessed through a proposed inter-class sample unbalance processing technology (including data enhancement based up-sampling processing and proportional generation batch method), so that the diversity of the training data is improved, and meanwhile, the training data sent into a network keeps balance among classes; through the proposed dense residual error network for SAR image ship target identification, the original characteristics can be reused while more new characteristics are learned; by the loss function based on the central loss, the parameters of the network model are adjusted, and the intra-class compactness and the inter-class separability are optimized simultaneously. Recognition results on OpenSARShip show that the designed dense residual network has higher accuracy, smaller model size and smaller calculation amount than the existing adopted neural network.

Corresponding to the SAR image ship target identification method, the invention also provides the following specific implementation system:

a SAR image ship target recognition system, comprising:

the training sample data set construction module is used for constructing a training sample data set by adopting a batch generation method according to a proportion;

the initial dense residual network acquisition module is used for acquiring an initial dense residual network;

the dense residual error network training module is used for training the initial dense residual error network by adopting the training sample data set to obtain a trained dense residual error network;

the SAR image data acquisition module is used for acquiring SAR image data to be identified;

the image data expansion module is used for performing data expansion according to the SAR image data in an image enhancement mode to obtain expanded image data;

and the ship target identification module is used for inputting the expanded image data into the trained dense residual error network to obtain a ship target identification result.

Preferably, the training sample data set constructing module specifically includes:

a ship type acquisition unit for acquiring a ship type;

the batch forming unit is used for selecting sample data with set quantity to form a batch based on each ship type;

and the training sample data set constructing unit is used for forming the training sample data set by the sample data of different batches.

Preferably, the batch forming unit specifically includes:

the sample processing subunit is used for counting the number of samples of each ship type and performing ascending arrangement according to the number of the samples;

the batch forming subunit is used for selecting a set number of samples in each category to form a batch; and when the residual number of the samples in the current ship category is smaller than the set number, selecting the residual samples and the samples at the beginning of the sample sequence to form a batch, and repeating the steps until the specified iteration times are reached.

Preferably, the dense residual network training module specifically includes:

a predicted value output unit, configured to input the training sample data set to the initial dense residual error network for feature extraction, and output a predicted value of a category;

the dense residual network training unit is used for determining a loss value according to a predicted value and a real value in the training sample data set based on a loss function of an initial dense residual network, updating model parameters of the initial dense residual network by using a gradient descent method based on the loss value until convergence training is finished, and obtaining a trained dense residual network; the loss function of the initial dense residual network is a loss function based on a central loss.

The technical effect achieved by the implementation system provided by the invention is the same as that achieved by the SAR image ship target identification method provided by the invention, so that the implementation system is not repeated herein.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

Fig. 1 is a flowchart of an SAR image ship target identification method provided by the present invention;

FIG. 2 is a flow chart of SAR image ship recognition training based on dense residual error network and central loss provided by the present invention;

FIG. 3 is a flow chart of SAR image ship identification based on dense residual error network and central loss provided by the present invention;

FIG. 4 is a schematic illustration of a scaled batch provided by the present invention;

FIG. 5 is a schematic diagram of a dense residual module provided by the present invention;

FIG. 6 is a block diagram of a dense residual error network DRNet-48 provided by the present invention;

fig. 7 is a schematic structural diagram of the SAR image ship target recognition system provided by the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

The invention aims to provide a method and a system for identifying SAR image ship targets, which have higher accuracy, smaller model size and smaller calculated amount.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

As shown in fig. 1, the method for identifying a ship target in an SAR image provided by the present invention comprises:

step 100: and constructing a training sample data set by adopting a batch generation method according to a proportion.

Step 101: an initial dense residual network is obtained. The initial dense residual error network combines the advantages of the dense module and the residual error module, and can realize the reutilization of the original characteristics while learning more new characteristics. The dense residual module can be regarded as an improvement on the dense module, a residual connection is added in the dense module to realize multiplexing of the previous layer features, and new features are learned through convolution operation on the basis.

Step 102: and training the initial dense residual error network by adopting a training sample data set to obtain a trained dense residual error network. The trained dense residual error network comprises a convolution layer, a pooling layer, a transformation layer, a classification layer and a plurality of dense residual error modules. The dense residual module is a feature learning module for aggregating the dense connection and the residual connection.

Step 103: and acquiring SAR image data to be identified.

Step 104: and performing data expansion according to the SAR image data in an image enhancement mode to obtain expanded image data. In particular, the invention performs upsampling based on data enhancement, and expands data in a data enhancement (rotation and cutting) mode, so that the richness of a data set can be improved compared with simple replication. The expanded data set will alleviate the sample imbalance problem to some extent.

Step 105: and inputting the expanded image data into the trained dense residual error network to obtain a ship target identification result.

Wherein, the number of each kind of samples entering the batch is ensured to be consistent, the above step 100 of the invention specifically comprises:

step 1001: and acquiring the ship category.

Step 1002: a set amount of sample data is selected based on each ship category to form a batch so that the samples within each batch are fully balanced. Specifically, 7 samples are selected per class (6 total) to form a batch based on the scaled-up batch. The specific operation process is as follows:

and counting the number of samples of each ship category, and performing ascending arrangement according to the number of the samples.

Selecting a set number of samples per category constitutes a batch. And when the residual number of the samples in the current ship category is smaller than the set number, selecting the residual samples and the samples at the beginning of the sample sequence to form a batch, and repeating the steps until the specified iteration times are reached.

Step 1003: and forming a training sample data set by different batches of sample data.

For example, in a specific application process, in a batch generated in proportion, serial numbers 0 to 6 of 6 types of targets are respectively selected (in a training data set, SAR image data of each type are respectively numbered from 0, and the serial numbers are the serial numbers of the data), so that a batch 1 is formed, serial numbers 7 to 12 of the 6 types of targets are selected, so that a batch 2 is formed, and the operation is repeated until a batch 36 is formed. In forming the batch 37, class 4 (dredge category) has only 2 samples (252 to 253), less than 7, and now 5 samples (0 to 4) are selected from the beginning of class 4, plus 252 and 253 to form the batch 37 for this training. The same strategy is adopted when other categories suffer from insufficient samples until the maximum category runs out of all samples.

Based on the above application process, the batch forming process may include 6 steps: step 1, counting the number of samples of each category, and rearranging the samples from small to large to obtain N1 to N6. And 2, dividing N1 to N6 by 7, rounding to obtain N1 to N6, wherein the corresponding remainders are m1 to m 6. Step 3, selecting 7 ship samples in each category to form a batch with the size of 42. Step 4, repeat step 3 until there is a class with a remainder less than 7, select the last remaining mi, and the first 7-mi to form batch 42. And 5, repeating the step 3 and the step 4 until the N1 has only m1 ships. Step 6, select the remaining m1 and the first 7-m1 to form a batch size of 42. Repeating the steps until reaching the specified iteration number and stopping.

Further, before obtaining the initial dense residual error network, the invention also includes constructing the dense residual error network. The dense residual error network combines the advantages of the dense module and the residual error module, and can realize the reutilization of the original characteristics while learning more new characteristics. The dense residual module can be regarded as an improvement on the dense module, a residual connection is added in the dense module to realize multiplexing of the previous layer features, and new features are learned through convolution operation on the basis.

Based on the constructed initial dense residual error network, training the initial dense residual error network by adopting a training sample data set to obtain a trained dense residual error network, which specifically comprises the following steps:

and inputting the training sample data set into an initial dense residual error network for feature extraction, and outputting a predicted value of the category.

And determining a loss value according to the predicted value and the true value in the training sample data set by using the loss function based on the initial dense residual error network, and updating the model parameters of the initial dense residual error network by using a gradient descent method based on the loss value until convergence training is finished to obtain the trained dense residual error network. The loss function of the initial dense residual network is a loss function based on the central loss.

Wherein the loss function based on the central loss is to increase the central loss on the basis of the softmax loss function. The softmax loss function is a loss function commonly used in the identification task, is easier to optimize, and can separate different classes of samples, namely, has stronger inter-class separability. The target identification in the SAR image is different from the target identification of images in other fields and mainly reflects the fact that the intra-class variation is large, because the SAR image of a ship is extremely sensitive to the observation angle, the appearance of the SAR image of the ship in the same class is greatly different due to slight variation of the angle, and therefore the loss function needs to use strong intra-class compactness. The new loss function can realize the simultaneous optimization of the compactness in the class and the separability between the classes, and realize the better convergence of the model.

Residual concatenation in the residual block combines the original input feature map with the feature maps subjected to convolution processing of 1 × 1, 3 × 3 and 1 × 1 using an element-by-element addition operation. The combination is essentially the multiplexing of the features in the previous layer, and has the advantages that the information of the previous layer in the network can be fused, and the feature utilization rate is high. And the dense module aggregates the current feature map with all the subsequent feature maps through dense connection and further convolves the aggregated features. Compared with the element-by-element addition operation in the residual error module, the aggregation operation is realized by splicing, the front-layer features are completely reserved, and more new features can be learned by the processing mode of 'aggregation and convolution' firstly. Meanwhile, compared with a residual error module, the dense module has stronger learning capacity for new features, but lacks the multiplexing of the previous features, and has the problem of low feature utilization rate.

The trained dense residual error network is preferably a dense residual error network DRNet-48, and the advantages of the dense module and the residual error module are combined, so that the original characteristics can be reused while more new characteristics are learned.

The dense residual network DRNet-48 structure is shown in table 1 below, the size of the input image is 1 × 32 × 32, the convolution kernel of the first convolution and pooling operation is large (16 × 16), and the reduction of information loss from the original input image can be avoided. The growth rate and decay factor in the table are 8 and 0.5, respectively. The 3 x 3 convolution in the table has a zero fill of 1 pixel, which leaves the feature map size unchanged after convolution. Each dense residual module is followed by a transform layer that includes 1 x 1 convolution and 2 x 2 average pooling. The 1 × 1 convolution is used to reduce the number of feature maps and prevent memory overload, and the 2 × 2 average pooling halves the feature map size. After the third dense residual module there will be transform layers without pooling used to reduce the number of channels of the feature map, classification layers used to classify the ships, and feature map sizes of the three dense residual modules are 8 × 8, 4 × 4, and 2 × 2, respectively.

TABLE 1 structural configuration Table of DRNet-48

The following describes a specific workflow of the SAR image ship target identification method provided by the present invention based on a specific embodiment.

The SAR image ship target identification method provided by the invention mainly comprises three improvements, which are respectively: inter-class sample imbalance processing techniques, dense residual networks, and loss functions based on center loss, as shown in phantom in fig. 2. The specific identification process is shown in fig. 3.

Wherein, 1. the sample imbalance among categories is processed

When training data is prepared, the data sent to the CNN model training is kept balanced as much as possible by adopting an inter-class sample imbalance processing technology, the method comprises two steps of data enhancement based up-sampling and proportional generation batch, the preprocessed data is input into the CNN network for model training, and the specific implementation process is as follows:

data enhancement based upsampling is an augmentation of data by means of data enhancement (rotation and cropping) that can increase the richness of a data set compared to simple replication. The expanded data set will alleviate the sample imbalance problem to some extent.

Aiming at the problem, an up-sampling strategy based on data enhancement is provided, and the method comprises the following steps:

the six categories of OpenSARShip, the maximum, are selected for classification experiments, cargo, tanker, fishing, tug, dredge, and "others", 1750, 491, 52, 26, 23, and 267, respectively.

The samples obtained above are subjected to data enhancement-based upsampling: the tanker is horizontally and vertically turned over to obtain 1473, the fishing boat, the tug boat and the dredge boat are respectively cut out on the upper left, the upper right, the lower left and the lower right, the numbers of 572, 286 and 253 are respectively obtained by horizontal and vertical turning, and the numbers of other types of fishing boats are respectively cut out on the upper left, the upper right, the lower left and the lower right to obtain 1602. Through the operation, the problem of unbalanced samples is relieved to a certain extent. The number of cargo, tanker, fishing, tug, dredge and "others" is 1750, 1473, 572, 286, 253 and 1602 respectively, and data enhancement increases the diversity of samples compared to simple replication.

After the training data is expanded, a batch generation method according to proportion is adopted in the batch selection process before the training model, so that the number of various samples entering the batch is consistent.

Fig. 4 is a schematic diagram of a batch generated to scale, and instead of randomly selecting 42 ship samples to form a batch, 7 samples are selected in each class (6 classes in total) to form a batch, so that the samples in each batch are fully balanced.

The sequence numbers of 0 to 6 of 6 types of targets are respectively selected in the batch generated according to the proportion (in a training data set, SAR image data of each type are respectively numbered from 0, the number is the sequence number of the data) to form a batch 1, the sequence numbers of 7 to 12 of the 6 types of targets are selected to form a batch 2, and the operation is repeated until the batch 36. In forming the batch 37, class 4 (dredge category) has only 2 samples (252 to 253), less than 7, and now 5 samples (0 to 4) are selected from the beginning of class 4, plus 252 and 253 to form the batch 37 for this training. The same strategy is adopted when other categories suffer from insufficient samples until the maximum category runs out of all samples.

The generation of batches in proportion comprises 6 steps: step 1, counting the number of samples of each category, and rearranging the samples from small to large to obtain N1-N6. And 2, dividing N1 to N6 by 7, rounding to obtain N1 to N6, wherein the corresponding remainders are m1 to m 6. Step 3, selecting 7 ship samples in each category to form a batch with the size of 42. Step 4, repeat step 3 until there is one type of remainder less than 7, select the last remaining mi, and the first 7-mi to form batch 42. And 5, repeating the

steps

3 and 4 until the N1 has only m1 ships. Step 6, select the remaining m1 and the first 7-m1 to form a batch size of 42. Repeating the steps until reaching the specified iteration number and stopping.

Through the above operations (scaling batches), although the data sets are unbalanced, the data in each batch is balanced, which makes the loss per time more accurate and the parameter update more accurate, so that better model parameters can be learned.

2. Dense residual error network

And taking the constructed batch as input to train the dense residual error network.

Residual concatenation in the residual block combines the original input feature map with the feature maps that have been convolved with 1 × 1, 3 × 3, and 1 × 1, using an element-by-element addition operation. The combination is essentially the multiplexing of the features in the previous layer, and has the advantages that the information of the previous layer in the network can be fused, and the feature utilization rate is high.

And the dense module aggregates the current feature map with all the subsequent feature maps through dense connection and further convolves the aggregated features. Compared with the element-by-element addition operation in the residual error module, the aggregation operation is realized by splicing, the front-layer features are completely reserved, and more new features can be learned by the processing mode of 'aggregation and convolution' firstly. Meanwhile, compared with a residual error module, the dense module has stronger learning capacity for new features, but lacks the multiplexing of the previous features, and has the problem of low feature utilization rate.

The invention combines the advantages of the two modules, designs the dense residual error module shown in figure 5 for processing the characteristics, and further realizes the reutilization of the original characteristics while learning more new characteristics. The dense residual module can be regarded as an improvement on the dense module, a residual connection is added in the dense module to realize multiplexing of the previous layer features, and new features are learned through convolution operation on the basis. On the basis of the dense residual module, a dense residual network DRNet-48 for SAR image ship target identification is constructed, and the main structure of the dense residual network DRNet-48 is shown in FIG. 6.

C, P and FC in fig. 6 represent convolution layer, pooling layer and full-link layer, respectively, after convolution of a large convolution kernel, the input SAR image will have 3 dense residual modules for processing, there will be a set of transform layers (with 1 × 1 convolution and 2 × 2 pooling configuration) between the dense residual modules, and finally the prediction categories (numbers of 1 to 6 represent 6 categories of targets, respectively) are output through a full-link layer. The dense residual module can recycle and learn new features by simultaneously introducing dense connection and residual connection into one module, thereby improving the feature expression capability and further improving the identification accuracy. The transform layer is used to reduce the feature map size by half the number of channels, which includes 1 × 1 convolution and 2 × 2 pooling. The transform layer without pooling is used to reduce the profile channel. The classification layer is used to output the final prediction classes, which represent 6 classes of objects with 6 numbers from 1 to 6, respectively. Meanwhile, DRNet-48 greatly reduces the number of channels, so that the model parameters and the calculated amount are greatly reduced, and the method is more suitable for SAR image ship target identification tasks. The detailed structural configuration of DRNet-48 is as shown in Table 1 above.

The proposed relationship of the network DRNet-48 to the networks DensenNet-48 and ResNet-50 is summarized as follows: the dense residual module in fig. 6 is replaced by a dense module to form the corresponding DenseNet-48. ResNet-50 refers to a 50-layer residual network constructed with multiple residual modules.

3. Loss function based on center loss

And the loss function based on the central loss is used for calculating the error of the predicted value and the true value at the output layer of the DRNet-48, and adjusting the model parameters of the DRNet-48 by a gradient descent method until convergence (the loss curve tends to be stable).

The invention adds a central loss function in the loss function, and the calculation formula is as follows:

in the formula, x_iRepresents the number of the i-th sample,

represents the y th_iCenter feature of individual class, for calculation

It is generally desirable to average the eigenvalues of each class in the batch before each iteration, with m representing the total number of samples.

Center loss function L₂The gradient of (a) is:

in the formula (I), the compound is shown in the specification,

the calculation formula of (c) is:

if y is_iJ, then δ (y)_iJ equals 1, whereas δ (y)_i＝j)＝0。

Center loss function L₂The model is enabled to have intra-class compactness, and a new loss function L (final loss function) is obtained by combining the softmax loss function with inter-class separability

L＝L₁+λL₂ (4)

In the formula, L₁Is the softmax loss, describes the change between classes, and separates the classes as much as possible. L is₂Is a central loss, representing intra-class variation, making the intra-class as compact as possible. λ is a constant used to balance the two types of losses, typically λ ═ 1. L is₁The calculation formula of (2) is as follows:

in the formula, x_iRepresents the number of the i-th sample,

and b_jRepresenting the weights and biases of the softmax layer neurons, respectively.

And (4) simultaneously calculating the central loss and the softmax loss, and accumulating the loss values obtained by the central loss and the softmax loss for adjusting the model parameters.

In the invention, the OpenSARShip data set is divided into a training set and a testing set according to the proportion of 8:2, and the batch size is 42. The learning rate from the 1 st to the 50 th training period is 0.1, the learning rate from the 51 st to the 100 th training period is 0.01, and the learning rate from the 101 st to the 150 th training period is 0.001.

The evaluation index is the accuracy P of averaging according to classes_classThe calculation formula is as follows:

in the formula, P_classMean accuracy, P, expressed over classes_class1To P_class6Representing the accuracy of class 6 targets, respectively.

In order to verify the specific influence of the inter-class sample imbalance processing technology and the loss function based on the central loss on the identification precision, the method adopts a control variable method, and trains and tests the SAR image in the SLC mode in the OpenSARShip data set by using different strategies on a ResNet-50 model.

Table 2 below is the test results of ResNet-50 trained on an unexpanded data set, with numbers 1 to 6 representing six categories, cargo, oil, fishing, tug, dredge, and "other," respectively. The accuracy of the class-wise averaging was 60.9%, and the accuracy of the sample-wise averaging was 458/521-87.9%. As can be seen from table 2, the identification effect of the rare classes is very poor, and the identification accuracy of the fishing boat, the tug boat and the dredge boat is not more than 50%.

TABLE 2 confusion matrix obtained with ResNet-50 on unexpanded data sets

The class-by-class accuracy of the training data after upsampling based on data enhancement was 64.6%. By up-sampling based on data enhancement, the accuracy of rare classes of fishing, tug and dredging vessels is improved from 50%, 40% and 20% to 51.6%, 50.9% and 35.2%, but the accuracy of other classes is still low. It can also be seen that as the rare class performance increases, the

classes

1, 2, 6 also decrease somewhat because the six classes of targets in the dataset have a competing role.

The confusion matrix of training data sorted after upsampling and scaled batch processing yields a class-wise averaging accuracy of 68.9%, which is a significant improvement over 60.9% and 64.7%, and scaling batches to scale up rare classes is more effective than other classes because the data sets are balanced within each batch despite the imbalance.

Table 3 is the experimental results of training data after enhanced upsampling based on data, scale-to-batch processing, and using the proposed loss function in ResNet-50, the accuracy of class-wise averaging is 73.5%. The improvement effect on all classes is obvious through the newly designed loss function, because the new loss function can simultaneously distinguish different classes and gather the same class, and the new loss function can be used as the loss function of a general classifier and is not limited to the unbalanced situation.

TABLE 3 confusion matrix obtained using the inter-class sample imbalance technique and the new loss function ResNet-48

To verify the recognition performance of the proposed DRNet-48 model, under the above settings, recognition experiments were performed using DenseNet-48 and DRNet-48, respectively. Under the same experimental conditions, DRNet-48 achieved the highest recognition accuracy (77.2%), and DenseNet achieved an accuracy of 74.9%, slightly higher than ResNet-50 (73.5%). The results of the comparative experiments show that DRNet-48, which is specifically designed for SAR images, has better feature extraction capability than ResNet-50 and DenseNet-50.

TABLE 4 confusion matrix obtained using the inter-class sample imbalance technique and the new loss function DRNet-48

For verification of identification performance, ResNet-50 and DenseNet-48 close to DRNet-48 of 48 layers were selected for comparative analysis. The model size generated for ResNet-50 is 97.75MB, the average iteration time to handle OpenSARShip is 23.55s, the model size for DRNet-48 is 15.73MB, the average iteration time to handle OpenSARShip is 5.27s, the model size for DenseNet-48 is 54.63MB, and the average iteration time to handle OpenSARShip is 21.34 s. The DRNet-48 has better classification performance, smaller model size and smaller calculation amount. The size and parameters of the model are much smaller (mainly the number of channels). This represents the superior performance of the model designed for SAR image target recognition by the present invention.

In addition, corresponding to the above provided method for recognizing the SAR image ship target, the present invention also provides a system for recognizing the SAR image ship target, as shown in fig. 7, the system includes: the system comprises a training sample data set construction module 1, an initial dense residual network acquisition module 2, a dense residual network training module 3, an SAR image data acquisition module 4, an image data expansion module 5 and a ship target identification module 6.

The training sample data set constructing module 1 is configured to construct a training sample data set by using a batch generation method according to a proportion.

The initial dense residual network obtaining module 2 is configured to obtain an initial dense residual network.

The dense residual network training module 3 is used for training the initial dense residual network by adopting a training sample data set to obtain a trained dense residual network.

The SAR image data acquisition module 4 is used for acquiring SAR image data to be identified.

The image data expansion module 5 is used for performing data expansion according to the SAR image data by adopting an image enhancement mode to obtain expanded image data.

And the ship target identification module 6 is used for inputting the expanded image data into the trained dense residual error network to obtain a ship target identification result.

Preferably, the training sample data set constructing module 1 specifically includes: the system comprises a ship type obtaining unit, a batch forming unit and a training sample data set constructing unit.

The ship type obtaining unit is used for obtaining the ship type.

The batch forming unit is used for selecting sample data with set quantity based on each ship category to form a batch.

The training sample data set construction unit is used for forming a training sample data set by different batches of sample data.

Preferably, the batch forming unit specifically includes: a sample processing subunit and a batch forming subunit.

The sample processing subunit is used for counting the number of samples of each ship category and performing ascending arrangement according to the number of the samples.

The batch forming subunit is used for selecting a set number of samples in each category to form a batch. And when the residual number of the samples in the current ship category is less than the set number, selecting the residual samples and the samples at the beginning of the sample sequence to form a batch, and repeating the steps until the specified iteration times are reached.

Preferably, the dense residual network training module specifically includes: a predicted value output unit and a dense residual error network training unit.

The predicted value output unit is used for inputting the training sample data set to the initial dense residual error network for feature extraction and outputting the predicted value of the category.

And the dense residual network training unit is used for determining a loss value according to the predicted value and a true value in the training sample data set based on a loss function of the initial dense residual network, updating model parameters of the initial dense residual network by using a gradient descent method based on the loss value until convergence training is finished, and obtaining the trained dense residual network. The loss function of the initial dense residual network is a loss function based on the central loss.

In the present specification, the embodiments are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.

The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the foregoing, the description is not to be taken in a limiting sense.

Claims

1. A SAR image ship target identification method is characterized by comprising the following steps:

acquiring an initial dense residual error network;

acquiring SAR image data to be identified;

inputting the expanded image data into the trained dense residual error network to obtain a ship target identification result;

the method for constructing the training sample data set by adopting the batch generation according to the proportion specifically comprises the following steps:

acquiring the class of a ship;

forming a training sample data set by different batches of sample data;

wherein, the selecting the set number of sample data based on each ship category forms a batch, which specifically includes:

2. The method for identifying a ship target according to SAR images of claim 1, wherein the training of the initial dense residual network with the training sample data set to obtain a trained dense residual network specifically comprises:

3. The SAR image ship target identification method of claim 1, wherein the trained dense residual network comprises a convolutional layer, a pooling layer, a transformation layer, a classification layer and a plurality of dense residual modules; the dense residual error module is a feature learning module for aggregating the dense connection and the residual error connection.

4. A SAR image ship target recognition system is characterized by comprising:

an initial dense residual network obtaining module, configured to obtain an initial dense residual network;

the ship target identification module is used for inputting the expanded image data into the trained dense residual error network to obtain a ship target identification result;

the training sample data set construction module specifically comprises:

a ship type acquisition unit for acquiring a ship type;

the training sample data set constructing unit is used for forming a training sample data set by different batches of sample data;

wherein, the batch forming unit specifically comprises:

the sample processing subunit is used for counting the number of samples of each ship category and performing ascending arrangement according to the number of the samples;

a batch forming subunit, configured to select a set number of samples in each category to form a batch; and when the residual number of the samples in the current ship category is smaller than the set number, selecting the residual samples and the samples at the beginning of the sample sequence to form a batch, and repeating the steps until the specified iteration times are reached.

5. The SAR image ship target recognition system of claim 4, wherein the dense residual network training module specifically comprises: