CN114882289A - SAR target open set identification method based on self-adaptive determination rejection criterion - Google Patents

Abstract

The invention discloses an SAR target open set identification method based on self-adaptive determination rejection criterion, which comprises the following implementation steps: generating a training set and a test set; constructing a convolution self-coding network and training; fitting the reconstruction error of the sample set by using an extreme value theory and obtaining the rejection threshold values of various targets; acquiring the prediction class probability of the SAR target to be identified by using a convolutional self-coding network; reconstructing an SAR target image to be identified and calculating a reconstruction error; and identifying the sample with the reconstruction error smaller than the rejection threshold corresponding to the prediction category as the category with the highest probability score of the prediction category, and identifying the sample with the reconstruction error larger than the rejection threshold as the unknown target. The method has the advantages of accurately identifying the known class and adaptively rejecting the unknown class, and solves the problem that the prior art cannot be used in engineering practice of a real open environment due to the lack of theoretical basis and poor generalization performance of rejection threshold setting.

Description

SAR target open set identification method based on self-adaptive determination rejection criterion

Technical Field

The invention belongs to the technical field of radar communication, and further relates to a Synthetic Aperture Radar (SAR) (synthetic Aperture radar) target open set identification method based on a self-adaptive determination rejection criterion in the technical field of automatic radar target identification. The method can be used for identifying the type of the SAR target in an open set scene under the radar interference environment.

Background

The radar Automatic target recognition technology rarr (radar Automatic target recognition) is a typical open-set scene recognition task, which is oriented to non-cooperative targets, and has numerous categories and various models, and it is difficult to establish a complete target recognition library in the training stage, i.e. there is an unknown target in the testing stage. The traditional pattern recognition technology is designed under the condition of a closed set assumption, namely, the set of class labels appearing in the training sample is assumed to contain the class labels of all samples to be recognized, and no unknown class target exists. In a real open environment, when an unknown class target outside a library enters a model, the closed set identification model can force the unknown class target to be mistakenly classified into a certain in-library category, so that the application of the model in the open environment is greatly limited. Therefore, a classifier capable of recognizing/rejecting an unknown class target while maintaining the performance of identifying the known class target is desired, in the radar automatic target recognition, a specific certain class is output for an input known class sample, and an unknown class is rejected for an input unknown class sample. The method comprises the steps of sensing unknown targets in the open environment, correctly and adaptively detecting the existence of the unknown targets and identifying the known targets, wherein the condition is the premise and the basis of identification in the radar open environment.

Ma Xiaojie et al, in its published paper "An Open Set Recognition Method for SAR Targets Based on Multitask Learning" (IEEE Geoscience and Remote Sensing Letters,2021,5), propose a Method for generating anti-network radar Open-Set Recognition in combination with classification branches. The method comprises a generator and a discriminator combined with a classification branch, wherein the generator inputs random noise and a class label and outputs a generated pseudo sample, and the discriminator combined with the classification branch inputs the pseudo sample and a real sample and outputs a corresponding discrimination score and a corresponding prediction label. The method comprises the following steps that firstly, random noise and a class label are input into a generator to obtain a generated image, and the generated image is mixed with a discriminator as much as possible; secondly, inputting the original image and the generated image into a discriminator combined with the classification branch to obtain corresponding confidence scores, wherein the discriminator is required to distinguish the original image and the generated image as much as possible; and thirdly, inputting the original image into a classification branch of the discriminator to minimize the cross entropy loss of the prediction label and the real label. Goodfellow compares the generator to the offender who prints counterfeit money and the discriminator to the policeman who catches the offender. Criminals forge genuine bills with false and misleading genuine bills, and policemen continuously improve the ability of identifying counterfeit bills and genuine bills. In the continuous game process of the generator and the discriminator, the discrimination capability of the discriminator and the generation capability of the generator are obviously improved. In the stage of setting the rejection criterion, after the model training is completed, the method traverses the confidence score distribution of all known class targets and unknown class targets on the discriminator to obtain the optimal threshold value. In the testing stage, a sample to be tested is input into a discriminator combined with the classification branch, corresponding confidence scores and prediction labels are output, when the discrimination score of the sample to be tested is smaller than the optimal threshold value obtained through traversal, the model rejects the sample to be tested as an unknown target, otherwise, the model identifies the sample to be tested as a category corresponding to the prediction labels. The method has the disadvantages that an unknown sample is required to be used for traversing and finding out a proper threshold value in the training process, and the unknown sample cannot be obtained in advance in engineering practice, so that the method has the defects that the judgment rejection threshold value selection strategy lacks a theoretical basis and is poor in generalization performance, and cannot be applied to the engineering practice in a real open environment.

The patent document "a radar high-resolution range profile open set identification method and device" (patent application No. CN202111199838.4, application publication No. CN 114137518 a) applied by the university of electronic science and technology of west ampere discloses a radar open set identification method. The method adopts the convolutional neural network technology to construct various types of prototype vectors to enhance intra-class cohesiveness, and simultaneously combines primary features of various layers, so that higher-layer features are obtained for recognition, and therefore, the recognition rate of the model on the known class target is remarkably improved. In the testing stage, when the classification score of the sample to be tested is smaller than the artificially set threshold value, the model rejects the sample to be tested as an unknown target, otherwise, the sample to be tested is identified as the category corresponding to the prediction label. The method has the disadvantages that potential characteristic information of unknown targets is not considered in the training stage, the performance of the model is very sensitive to a rejection threshold for removing the unknown targets, the artificially set rejection threshold lacks theoretical basis and has poor generalization performance, and when the method is directly applied to engineering practice in a real open environment, the robust performance of an SAR target open set identification model is weak, and the unknown targets are difficult to accurately reject and identify the known targets.

Disclosure of Invention

The invention aims to provide an SAR target open set identification technology based on a self-adaptive determination rejection criterion aiming at overcoming the defects of the prior art, and aims to solve the problem that the prior art is not suitable for engineering practice of a real open environment due to the lack of theoretical basis and poor generalization performance of the rejection threshold setting.

The idea for realizing the purpose of the invention is as follows: on the premise of only using known training samples, the method is characterized in that the reconstruction error distribution of unknown targets is fitted in a training stage, the reconstruction errors are modeled through a statistical modeling extreme value theory, the rejection threshold is further determined in a self-adaptive mode, rejection of the known targets in the library and the unknown targets out of the library is achieved, and the problem that the method cannot be applied to engineering practice in a real open environment due to the fact that the rejection threshold is lack of theoretical basis in the prior art is solved. The method designs a convolution self-coding model combined with a channel level attention mechanism for an SAR target open set identification task by combining with SAR target characteristics, extracts joint optimal characteristic representation of an identification module and a reconstruction module by a joint optimization strategy, further improves the identification rate of the model to the SAR target, and solves the problem that the prior art cannot be applied to engineering practice in a real open environment due to poor generalization performance of the model.

The technical scheme adopted by the invention comprises the following steps:

step 1, generating a training set:

selecting at least 1000 SAR images with a radar working pitch angle of 17 degrees, cutting and labeling each SAR image, and forming a training set by all cut images and corresponding class labels;

step 2, constructing an open set identification model of the convolutional self-coding model:

step 2.1, constructing a coding sub-network:

building a coding sub-network formed by connecting five channel attention modules with the same structure in series;

each channel attention module consists of two parallel branches:

the structure of the first branch is as follows in sequence: a convolutional layer, a Batch Norm normalization layer, a ReLU nonlinear layer; the number of convolution kernels of a first branch convolution layer in the first module to the fifth module is set to be 16, 32, 64, 128 and 256 in sequence, the sizes of the convolution kernels are all set to be 3 multiplied by 3, the step sizes of the convolution kernels are all set to be 1, and the filling modes are all set to be equal-size filling modes;

the structure of the second branch is as follows in sequence: the global average pooling layer, the first convolution layer, the ReLU nonlinear layer, the second convolution layer and the Sigmoid activation function; sequentially setting the number of convolution kernels of a first convolution layer in a second branch circuit in the first module, the second module, the third module and the fourth module to be 4, 8, 16, 32 and 64, setting the sizes of the convolution kernels to be 1 multiplied by 1, setting the step length of the convolution kernels to be 1, and setting the filling modes to be equal-size filling modes; the number of convolution kernels of a second convolution layer of a second branch in the first module to the fifth module is set to be 16, 32, 64, 128 and 256 in sequence, the sizes of the convolution kernels are all set to be 1 multiplied by 1, the step length of the convolution kernels is all set to be 1, and the filling modes are all set to be equal-size filling modes;

step 2.2, constructing a conditional sub-network:

building a conditional subnetwork consisting of three fully-connected layers; setting the node number of the full connection layer as 256, 1024 and 4096 in sequence;

step 2.3, constructing a classification sub-network:

building a classification sub-network consisting of two convolutional layers; the number of convolution kernels is set to 128 and K in sequence; wherein K represents a training set category; setting the sizes of convolution kernels to be 1 multiplied by 1, setting the step lengths of the convolution kernels to be 1, and setting the filling modes to be equal-size filling modes;

step 2.4, constructing a decoding sub-network:

building a decoding sub-network formed by connecting four deconvolution layers in series; the number of convolution kernels of the first to fourth deconvolution layers is set to be 128, 64, 32 and 16 in sequence, the sizes of the convolution kernels are all set to be 4 multiplied by 4, the step sizes of the convolution kernels are all set to be 2, the filling modes are all set to be equal-size filling modes, and the deviation positions are all set to be 0;

step 2.5, connecting the condition sub-network and the coding sub-network in parallel to obtain a sub-network 1, connecting the classification sub-network and the decoding sub-network in parallel to obtain a sub-network 2, and connecting the sub-network 1 and the sub-network 2 in series to form an open set identification model based on convolution self-coding;

step 3, training an open set recognition model:

step 3.1, inputting each image of the training set into an open set identification model, outputting the sample hidden characteristics of each image in the training set after passing through a coding sub-network, and outputting the prediction class probability of each image after passing through a classification sub-network; calculating the cross entropy loss L between the prediction category probability of each image and the category label corresponding to the image by using a cross entropy loss function _c ：

Step 3.2, inputting each image category label of the training set into the open set identification model, and outputting the conditional implicit characteristic of each image through a conditional sub-network; multiplying the sample hidden feature of each image with the corresponding conditional hidden feature point to obtain the matching fusion feature of the image; matching of matching fusion features output per image via decoded sub-networkPreparing a reconstructed image; multiplying the sample hidden feature of each image with the condition hidden feature point which does not correspond to the sample hidden feature to obtain the non-matching fusion feature of the image; outputting a non-matching reconstructed image of each image by the decoding sub-network according to the non-matching fusion characteristics; calculating the reconstruction error between the matched reconstructed image and the image, and calculating the reconstruction error between the non-matched reconstructed image and the randomly selected other images which are not the image category to obtain the reconstruction loss L _d ：

Step 3.3, summing the cross entropy loss obtained by calculation in the step 3.1 and the reconstruction loss obtained by calculation in the step 3.2 to obtain the total loss of the model; iteratively updating the open set identification model parameters by using an Adam optimization algorithm until the total loss function is converged to obtain a trained open set identification model;

step 4, adaptively setting a rejection threshold value of each category by using an extreme value theory:

step 4.1, inputting each image and category label in the training set into the trained open set identification model, and outputting a matched reconstructed image and a non-matched reconstructed image of each image; calculating matching reconstruction errors and non-matching reconstruction errors of each category of the training set;

step 4.2, respectively carrying out extreme value distribution fitting on the matching reconstruction errors and the non-matching reconstruction errors of each category of the training set based on an extreme value theory, and outputting the extreme value distribution of each category of the training set;

step 4.3, traversing the reconstruction errors of each category in the training set to obtain the extreme value distribution minimum value of each category, and taking the reconstruction errors corresponding to the extreme value distribution minimum value of each category as the optimal rejection threshold value of each category;

step 5, obtaining the prediction category probability and the reconstructed image of the SAR image to be identified:

cutting the SAR image to be recognized in the same way as the step 1 to obtain a cut SAR image, inputting the cut SAR image into a trained convolutional self-coding network, outputting the prediction class probability of the SAR image and a corresponding reconstructed image, and taking the class corresponding to the highest score value in the prediction class probability as the prediction class of the SAR image to be recognized;

step 6, judging whether the reconstruction error of the SAR image to be identified is smaller than the optimal judgment rejection threshold value of the prediction type, if so, executing step 7, otherwise, executing step 8;

step 7, selecting the category with the highest probability score of the SAR image prediction category as an identification result to be output;

and 8, judging the SAR image as an unknown target and outputting the unknown target.

Compared with the prior art, the invention has the following advantages:

firstly, on the premise of only using a known training sample, the method fits the reconstruction error distribution of the unknown target and models the reconstruction error through the extreme value theory of statistical modeling, so as to adaptively determine the rejection threshold value, realize the rejection of the known target in the library and the unknown target out of the library, and overcome the problem that the rejection accuracy of the model to the unknown target is rapidly reduced when the rejection threshold value setting lacks theoretical basis and the generalization performance is poor in the prior art, so that the method has the advantages of accurately identifying the known target and efficiently detecting the unknown target.

Secondly, the method considers the SAR image characteristics and the balance of the model performance and parameter quantity, designs a convolution self-coding model combined with a channel level attention mechanism, is used for an SAR target open set identification task, and extracts the joint optimal characteristic representation of an identification module and a reconstruction module through a joint optimization strategy, so that the method has the advantages of automatically extracting the characteristics of the SAR target image, enabling the identification reconstruction process to be an end-to-end process, being high in real-time performance, enabling the training speed of the model to be fast and not prone to overfitting through a batch normalization layer.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

FIG. 2 is a schematic view of an overall model of the present invention;

FIG. 3 is a model schematic of the attention convolution module of the present invention;

FIG. 4 is a graph comparing the open set performance indicator results of the present invention with the prior art.

Detailed Description

The invention is described in further detail below with reference to the figures and specific examples.

The implementation steps of the present invention are described in further detail with reference to fig. 1 and a specific embodiment.

Step 1, generating a training set and a testing set.

Step 1.1, selecting 5172 images from an SAR vehicle target data set MSTAR (moving and static target acquisition) to form a sample set;

step 1.2, cutting each image of 2746 images with a working pitch angle of 17 degrees in a sample set into images of 64 multiplied by 64 pixels, wherein the class label of each image is 1-10, and then forming a training set by all the cut images and the corresponding class labels;

step 1.3, cutting each image in 2426 images with the radar working pitch angle of 15 degrees in the sample set into images with 64 multiplied by 64 pixels, wherein the class label of each image belongs to 1-10, and then forming a test set by all the cut images and the corresponding class labels;

step 2, constructing an open set identification model of the convolutional self-coding model:

the coding sub-network, the decoding sub-network, the classification sub-network and the condition sub-network of the constructed open set identification model of the convolutional self-coding network are described in detail with reference to fig. 2.

Step 2.1, constructing a coding sub-network:

a coding sub-network consisting of five channel attention modules with the same structure connected in series is constructed, as shown in fig. 3. And inputting an original image into the coding sub-network, and outputting a sample hidden feature obtained by feature extraction.

Each channel attention module consists of two parallel branches:

the structure of the first branch is as follows in sequence: convolutional layer, Batch Norm normalization layer, ReLU nonlinear layer. The number of convolution kernels of a first branch convolution layer in the first module, the second module, the third module, the fourth module and the fifth module is set to be 16, 32, 64, 128 and 256 in sequence, the sizes of the convolution kernels are all set to be 3 multiplied by 3, the step lengths of the convolution kernels are all set to be 1, and the filling modes are all set to be equal-size filling modes;

the structure of the second branch is as follows in sequence: the device comprises a global average pooling layer, a first convolution layer, a ReLU nonlinear layer, a second convolution layer and a Sigmoid activation function. The number of convolution kernels of a first convolution layer in a second branch circuit in the first module to the fifth module is set to be 4, 8, 16, 32 and 64 in sequence, the sizes of the convolution kernels are all set to be 1 multiplied by 1, the step length of the convolution kernels is all set to be 1, and the filling modes are all set to be equal-size filling modes. The number of convolution kernels of the second convolution layer of the second branch in the first module to the fifth module is set to be 16, 32, 64, 128 and 256 in sequence, the sizes of the convolution kernels are all set to be 1 multiplied by 1, the step sizes of the convolution kernels are all set to be 1, and the filling modes are all set to be equal-size filling modes.

The coding sub-network adopts the channel attention module to carry out channel-level weighting on the feature map obtained by each convolution layer, establishes the interdependency relation among the feature channels, can enable the model to automatically learn the importance of each feature channel in the feature map, highlights the important features and inhibits useless features;

step 2.2, constructing a conditional sub-network:

and constructing a conditional subnetwork consisting of three fully-connected layers. Setting the node number of the full connection layer as 256, 1024 and 4096 in sequence; inputting a category label into the conditional sub-network, and outputting a conditional implicit feature obtained by feature extraction;

step 2.3, constructing a classification sub-network:

a classification subnetwork consisting of two convolutional layers is constructed. The number of convolution kernels is set to 128, K in order. Where K represents a training set class. The sizes of convolution kernels are all set to be 1 multiplied by 1, the step lengths of the convolution kernels are all set to be 1, and the filling modes are all set to be equal-size filling modes; the classification sub-network inputs the sample hidden characteristics extracted by the coding sub-network and outputs the prediction class probability, and the class corresponding to the maximum probability value is a prediction label;

step 2.4, constructing a decoding subnetwork:

and building a decoding sub-network consisting of four deconvolution layers. The number of convolution kernels is set to be 128, 64, 32 and 16 in sequence, the sizes of the convolution kernels are all set to be 4 multiplied by 4, the step sizes of the convolution kernels are all set to be 2, the filling modes are all set to be equal-size filling modes, and the deviation positions are all set to be 0; decoding the fusion characteristics multiplied by the hidden characteristics of the input sample of the sub-network and the hidden characteristics of the condition, and outputting a reconstructed image;

and 2.5, connecting the condition sub-network and the coding sub-network in parallel to obtain a sub-network 1, connecting the classification sub-network and the decoding sub-network in parallel to obtain a sub-network 2, and connecting the sub-network 1 and the sub-network 2 in series to form an open set identification model based on convolutional self-coding.

Step 3, training an open set recognition model:

and 3.1, inputting each image of the training set into the open set identification model, outputting the sample hidden characteristics of each image in the training set after passing through a coding sub-network, and outputting the prediction class probability of each image after passing through a classification sub-network. Calculating a loss value between the prediction category probability of each image and the category label corresponding to the image by using a cross entropy loss function:

where J represents the cross entropy loss function, N represents the total number of images in the training set, Sigma represents the summation operation, i represents the number of images in the training set, Y represents the number of images in the training set _i Representing the class label corresponding to the ith image in the training set, log representing the base 2 logarithm operation, P _i Representing the prediction category probability obtained by inputting the ith image in the training set into the open set recognition model;

and 3.2, inputting each image category label of the training set into the open set recognition model, and outputting the conditional implicit characteristics of each image through a conditional sub-network. And multiplying the sample hidden feature of each image by the corresponding conditional hidden feature point to obtain the matched fusion feature of the image. Matching the fusion characteristics and outputting a matching reconstruction image of each image through a decoding sub-network; and multiplying the sample hidden features of each image by the condition hidden feature points which do not correspond to the sample hidden features to obtain the non-matching fusion features of the image. The non-matching fused features output a non-matching reconstructed image for each image via the decoding sub-network. Calculating the reconstruction error between the matched reconstructed image and the image, and calculating the non-matched reconstructed imageThe reconstruction error with other randomly selected image types which are not the image type is obtained to obtain the reconstruction loss value L _d ：

Wherein, λ represents weight parameter in loss function, and its value range is [0,1]The larger the parameter value is, the stronger the reconstruction capability of the model for the matching reconstructed image is, and the weaker the reconstruction capability for the non-matching reconstructed image is. Considering that the open set identification model of the invention needs to have stronger reconstruction capability on the matching reconstructed image, λ is set to be a larger value of 0.8 in the embodiment of the invention. I | · | purple wind ₁ Representing a 1-norm operation.

And

respectively representing a matching reconstructed image and a non-matching reconstructed image of the ith image.

And 3.3, summing the cross entropy loss obtained by calculation in the step 3.1 and the reconstruction loss value obtained by calculation in the step 3.2 to obtain a model total loss function. And using an Adam optimization algorithm to iteratively update the open set identification model parameters until the total loss function is converged to obtain a trained open set identification model.

In the embodiment of the invention, all training processes use an Adam optimizer, the Adam first-order exponential decay factor is 0.9, and the second-order exponential decay factor is 0.999;

step 4, adaptively setting a rejection threshold value of each category by using an extreme value theory:

and 4.1, after model training is finished, inputting each image and category label in the training set into the open set identification model, and finally obtaining a matched reconstructed image and a non-matched reconstructed image of each image. Calculating the probability density distribution of the matching reconstruction errors and the probability density distribution of the non-matching reconstruction errors of each category of the training set;

and 4.2, respectively carrying out extreme value distribution fitting on the matching reconstruction error probability density distribution and the non-matching reconstruction error probability density distribution of each category of the training set based on an extreme value theory to obtain the extreme value distribution of each category of the training set. In the embodiment of the invention, the extreme value Distribution is set as generalized Pareto Distribution (Pareto Distribution);

and 4.3, based on the extreme value theory and the constructed training strategy, fitting to obtain extreme value distribution in a convex function form, and having an explicit optimal solution. Traversing the reconstruction error of each category in the training set to obtain an extreme value distribution minimum value of each category, and taking the reconstruction error corresponding to the extreme value distribution minimum value of each category as the optimal rejection threshold value of each category;

step 5, testing the open set identification model:

step 5.1, inputting each image in the test set into a trained open set identification model, outputting the sample hidden characteristics of each image in the test set through a coding sub-network, and outputting the prediction category probability and the corresponding prediction label of each image through a classification sub-network;

and 5.2, inputting each image prediction label in the test set into the open set identification model, and outputting the conditional implicit characteristics of each image in the test set through a conditional sub-network. And multiplying the sample hidden feature of each image by the corresponding conditional hidden feature point to obtain the matched fusion feature of the image. And outputting the matched reconstructed image of each image by the matched fusion characteristics through a decoding sub-network, and calculating the reconstruction error of the matched reconstructed image and the original image. If the reconstruction error is larger than the rejection threshold value set by the prediction label corresponding to the class self-adaption, the sample to be detected is judged as an out-of-library unknown class target, otherwise, the sample to be detected is identified as the pre-prediction output by the classifier subnetwork

And (6) detecting the label.

The effect of the present invention will be further described with reference to simulation experiments.

1. Simulation conditions are as follows:

the hardware platform of the simulation experiment of the invention is as follows: the processor is an Intel i 910900 CPU, the main frequency is 2.80GHz, and the memory is 16 GB.

The software platform of the simulation experiment of the invention is as follows: the Windows 10 operating system, MATLAB R2021a, and python 3.7.

The data set used by the simulation experiment is an SAR vehicle target data set MSTAR (moving and static target acquisition), the data in the data set is high-resolution bundled synthetic aperture radar which collects multiple military vehicles with Soviet Union targets, the imaging time is in the middle of the nineties of the twentieth century, the image size is 128 multiplied by 128 pixels, the image contains SAR image data of seven targets such as 9 different pitch angles about BTR60, 2S1 and BRDM2, and the image format is mat. In the simulation experiment, all images with the pitch angle of 17 degrees in a data set form a training sample set, and 2746 images are provided in total. All images in the data set with a 15-degree pitch angle form a test sample set, and 2426 images are obtained in total.

2. Simulation content and result analysis thereof:

the simulation experiment of the invention is to adopt the invention and two prior arts (conditional self-encoding C2AE open-set identification method, convolutional neural network OpenMax open-set identification method) to respectively carry out open-set identification on the input military vehicle SAR image.

In the simulation experiment, two prior arts are adopted:

the prior art conditional self-encoding C2AE Open set identification method refers to an optical image Open set identification method, called C2AE Open set identification method for short, which is proposed by Poojan Oza et al in ' C2AE: Class Conditioned Auto-Encoder for Open-set registration, IEEE Conference on Computer Vision and Pattern Registration (CVPR) ' 2019 '.

The OpenMax open set identification method of the convolutional neural network in the prior art refers to an optical image open set identification method, which is put forward in "optics set deep networks, IEEE Conference on Computer Vision and Pattern Registration (CVPR). 2016" by Bendale a et al, and is called OpenMax open set identification method for short.

The simulation experiment of the invention is to randomly select K classes and corresponding class labels from a training sample set to form a training set (K value range is [3,7 ]). All ten military vehicle images and category labels with a radar operating pitch angle of 15 ° are used to form a test set. 5 repeated open set identification simulation experiments were performed at each openness.

In order to ensure fairness, the model framework of the method and the comparison method are consistent, and the division of the known class/unknown class data sets of the repeated experiments is the same.

And (3) evaluating the open set identification results of the invention and the comparison method under different openness degrees by utilizing an evaluation index (F1 score). F1 score was calculated using the following formula, all calculations plotted in fig. 4:

wherein, TP _i Indicating the number of correctly identified images of the i-th class, FN _i Number of images, FP, representing false rejection of class i images _i Indicating the number of pictures, TN, for which the unknown class is predicted as class i _i Indicating the number of images for which the unknown class was correctly rejected. K denotes a training set class.

The effect of the simulation experiment of the present invention is further described with reference to fig. 4.

C2AE and OpenMax in fig. 4 represent two automatic threshold determination methods in the prior art, respectively, and the open set performance index F1 score of the experiment is repeated at random for multiple times under different openness degrees. The Proposed broken line in fig. 4 represents the open set identification technology Proposed by the present invention, and the open set performance index F1 score of the experiment is randomly repeated many times under different openness. The invention can keep higher F1 score index under different openness degrees, and the Proposed technology is improved by 3 percent compared with the prior art under the condition of small openness degree; in case of large openness, the present invention is more robust than the prior art, and F1 score indicates 5 percentage points. As can be seen from fig. 4, the SAR target open set identification technology for adaptively determining the rejection criterion provided by the present invention is more robust against a real open environment.

The foregoing description is only an example of the present invention and should not be construed as limiting the invention in any way, and it will be apparent to those skilled in the art that various changes and modifications in form and detail may be made therein without departing from the principles and arrangements of the invention, but such changes and modifications are within the scope of the invention as defined by the appended claims.

Claims (3)

1. A SAR target open set identification method based on self-adaptive determination rejection criterion is characterized in that a convolution self-coding network for outputting prediction category probability of an SAR target is constructed, and whether the SAR target is an unknown target is detected by self-adaptively setting the rejection threshold criterion; the identification method comprises the following steps:

step 1, generating a training set:

selecting at least 1000 SAR images with a radar working pitch angle of 17 degrees, cutting and labeling each SAR image, and forming a training set by all cut images and corresponding class labels;

step 2, constructing an open set identification model of the convolutional self-coding model:

step 2.1, constructing a coding sub-network:

building a coding sub-network formed by connecting five channel attention modules with the same structure in series;

each channel attention module consists of two parallel branches:

the structure of the first branch is as follows in sequence: a convolutional layer, a Batch Norm normalization layer, a ReLU nonlinear layer; the number of convolution kernels of a first branch convolution layer in the first module to the fifth module is set to be 16, 32, 64, 128 and 256 in sequence, the sizes of the convolution kernels are all set to be 3 multiplied by 3, the step sizes of the convolution kernels are all set to be 1, and the filling modes are all set to be equal-size filling modes;

the structure of the second branch is as follows in sequence: the global average pooling layer, the first convolution layer, the ReLU nonlinear layer, the second convolution layer and the Sigmoid activation function; sequentially setting the number of convolution kernels of a first convolution layer in a second branch circuit in the first module, the second module, the third module and the fourth module to be 4, 8, 16, 32 and 64, setting the sizes of the convolution kernels to be 1 multiplied by 1, setting the step length of the convolution kernels to be 1, and setting the filling modes to be equal-size filling modes; the number of convolution kernels of a second convolution layer of a second branch in the first module to the fifth module is set to be 16, 32, 64, 128 and 256 in sequence, the sizes of the convolution kernels are all set to be 1 multiplied by 1, the step length of the convolution kernels is all set to be 1, and the filling modes are all set to be equal-size filling modes;

step 2.2, constructing a conditional sub-network:

building a conditional subnetwork consisting of three fully-connected layers; setting the node number of the full connection layer as 256, 1024 and 4096 in sequence;

step 2.3, constructing a classification sub-network:

building a classification sub-network consisting of two convolutional layers; the number of convolution kernels is set to 128 and K in sequence; wherein K represents a training set category; setting the sizes of convolution kernels to be 1 multiplied by 1, setting the step lengths of the convolution kernels to be 1, and setting the filling modes to be equal-size filling modes;

step 2.4, constructing a decoding sub-network:

building a decoding sub-network formed by connecting four deconvolution layers in series; the number of convolution kernels of the first to fourth deconvolution layers is set to be 128, 64, 32 and 16 in sequence, the sizes of the convolution kernels are all set to be 4 multiplied by 4, the step sizes of the convolution kernels are all set to be 2, the filling modes are all set to be equal-size filling modes, and the deviation positions are all set to be 0;

step 2.5, connecting the condition sub-network and the coding sub-network in parallel to obtain a sub-network 1, connecting the classification sub-network and the decoding sub-network in parallel to obtain a sub-network 2, and connecting the sub-network 1 and the sub-network 2 in series to form an open set identification model based on convolution self-coding;

step 3, training an open set recognition model:

step 3.1, inputting each image of the training set into an open set identification model, outputting the sample hidden characteristics of each image in the training set after passing through a coding sub-network, and outputting the prediction class probability of each image after passing through a classification sub-network; by using the cross-entropy loss function,calculating the cross entropy loss L between the prediction class probability of each image and the class label corresponding to the image _c ：

Step 3.2, inputting each image category label of the training set into the open set identification model, and outputting the conditional implicit characteristic of each image through a conditional sub-network; multiplying the sample hidden feature of each image with the corresponding conditional hidden feature point to obtain the matching fusion feature of the image; matching the fusion characteristics and outputting a matching reconstruction image of each image through a decoding sub-network; multiplying the sample hidden feature of each image with the condition hidden feature point which does not correspond to the sample hidden feature to obtain the non-matching fusion feature of the image; outputting a non-matching reconstructed image of each image by the decoding sub-network according to the non-matching fusion characteristics; calculating the reconstruction error of the matched reconstructed image and the image, calculating the reconstruction error of the non-matched reconstructed image and the randomly selected other images which are not the image category, and obtaining the reconstruction loss L _d ：

Step 3.3, summing the cross entropy loss obtained by calculation in the step 3.1 and the reconstruction loss obtained by calculation in the step 3.2 to obtain the total loss of the model; iteratively updating the open set identification model parameters by using an Adam optimization algorithm until the total loss function is converged to obtain a trained open set identification model;

step 4, adaptively setting a rejection threshold value of each category by using an extreme value theory:

step 4.1, inputting each image and category label in the training set into the trained open set identification model, and outputting a matched reconstructed image and a non-matched reconstructed image of each image; calculating matching reconstruction errors and non-matching reconstruction errors of each category of the training set;

step 4.2, respectively carrying out extreme value distribution fitting on the matching reconstruction errors and the non-matching reconstruction errors of each category of the training set based on an extreme value theory, and outputting the extreme value distribution of each category of the training set;

step 4.3, traversing the reconstruction errors of each category in the training set to obtain the extreme value distribution minimum value of each category, and taking the reconstruction errors corresponding to the extreme value distribution minimum value of each category as the optimal rejection threshold value of each category;

step 5, obtaining the prediction category probability and the reconstructed image of the SAR image to be identified:

cutting the SAR image to be recognized in the same way as the step 1 to obtain a cut SAR image, inputting the cut SAR image into a trained convolutional self-coding network, outputting the prediction class probability of the SAR image and a corresponding reconstructed image, and taking the class corresponding to the highest score value in the prediction class probability as the prediction class of the SAR image to be recognized;

step 6, judging whether the reconstruction error of the SAR image to be identified is smaller than the optimal judgment rejection threshold value of the prediction type, if so, executing step 7, otherwise, executing step 8;

step 7, selecting the category with the highest probability score of the SAR image prediction category as an identification result to be output;

and 8, judging the SAR image as an unknown target and outputting the unknown target.

2. The SAR target open set identification method based on the self-adaptive determination rejection criterion as claimed in claim 1, characterized in that the cross entropy loss L in step 3.1 _c The following were used:

where N denotes the total number of images in the training set, Sigma denotes the summing operation, i denotes the number of images in the training set, Y _i Representing the class label corresponding to the ith image in the training set, log representing a base 2 logarithm operation, P _i And the prediction category probability obtained by inputting the ith image in the training set into the open set recognition model is shown.

3. The SAR target open set identification method based on the adaptive determination rejection criterion according to claim 1, characterized in that the reconstruction loss L in step 3.2 _d The following were used:

wherein, λ represents weight parameter in loss function, and its value range is [0,1]The larger the parameter value is, the stronger the reconstruction capability of the model on the matched reconstructed image is, and the weaker the reconstruction capability on the non-matched reconstructed image is; i | · | live through ₁ Representing 1-norm operation corresponding to reconstruction error calculation;

and

respectively representing a matching reconstructed image and a non-matching reconstructed image of the ith image.

Images