CN116030300A - Progressive domain self-adaptive recognition method for zero-sample SAR target recognition - Google Patents
Progressive domain self-adaptive recognition method for zero-sample SAR target recognition Download PDFInfo
- Publication number
- CN116030300A CN116030300A CN202211711796.2A CN202211711796A CN116030300A CN 116030300 A CN116030300 A CN 116030300A CN 202211711796 A CN202211711796 A CN 202211711796A CN 116030300 A CN116030300 A CN 116030300A
- Authority
- CN
- China
- Prior art keywords
- pseudo
- layer
- measured data
- label
- actual measurement
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 57
- 230000000750 progressive effect Effects 0.000 title claims abstract description 33
- 238000004088 simulation Methods 0.000 claims abstract description 80
- 238000005259 measurement Methods 0.000 claims abstract description 56
- 238000012545 processing Methods 0.000 claims abstract description 7
- 230000003044 adaptive effect Effects 0.000 claims description 26
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000012549 training Methods 0.000 description 53
- 238000002474 experimental method Methods 0.000 description 26
- 230000006870 function Effects 0.000 description 23
- 230000008569 process Effects 0.000 description 8
- 238000012360 testing method Methods 0.000 description 7
- 230000006978 adaptation Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 238000011524 similarity measure Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000003416 augmentation Effects 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013501 data transformation Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002779 inactivation Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000009966 trimming Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a progressive domain self-adaptive identification method for zero-sample SAR target identification, which comprises the following steps: importing the simulation data sample and the actually measured data sample into a feature alignment network, and extracting first feature information and second feature information; calculating according to the first characteristic information and the second characteristic information to obtain distribution difference; obtaining a trained feature alignment network based on the first loss function; predicting a label-free actual measurement data set by utilizing the trained characteristic alignment network, and endowing a pseudo label to the label-free actual measurement data set to obtain a first actual measurement data set of the pseudo label; performing pseudo-tag denoising processing on the first measured data set to obtain a pseudo-tag denoising measured data set; introducing the actual measurement data set after the pseudo tag denoising into the trained characteristic alignment network to adjust network parameters; and inputting the data to be classified into the adjusted characteristic alignment network to obtain a final classification and identification result. The invention effectively removes the pseudo tag data containing noise, thereby enabling the recognition result to be more stable.
Description
Technical Field
The invention belongs to the technical field of radars, and particularly relates to a progressive domain self-adaptive recognition method for zero-sample SAR target recognition.
Background
Synthetic Aperture Radar (SAR) data has been widely used in many fields such as military reconnaissance, disaster early warning, environmental monitoring, etc. The SAR Automatic Target Recognition (ATR) technique, which aims at recognizing the target type of the SAR image, has been widely studied. For SAR-ATR, a large amount of tag data is typically required to train the classifier in order to ensure efficient recognition performance of the classifier. Thus, the lack of training data will affect the recognition performance. The generation of SAR images by simulation techniques is an effective way to acquire data. However, for various reasons, there may be a difference between the distribution of the simulation image and the measured image, resulting in limitations in the application of the simulation image. Therefore, the key to using the simulation image is to overcome the difference between the simulation image and the measured image.
In the last few years, the SAMPLE data set provided by the united states Air Force Research Laboratory (AFRL), which contains simulated SAR target data, was used for the study of zero SAMPLE SAR target identification problems.
Some scholars have studied the application of simulation data in recognition, and the application can be roughly divided into the following categories:
(1) Pre-training.
Mallgren-Hansen et al in 2017 published paper "Improving SAR Automatic Target Recognition Models With Transfer Learning From Simulated Data" (IEEE Geoscience and Remote Sensing Letters) pre-trains the classifier with a large amount of simulated SAR data, and then fine-tunes the pre-trained network with a small amount of measured SAR data.
(2) Data transformation or data enhancement.
Liu et al in 2018, paper "SAR Target Classification with CycleGAN Transferred Simulated Samples" (IEEE International Geoscience & Remote Sensing Symposium) proposed the use of a cyclic consistency generation countermeasure network (CycleGAN) to transform simulation data to make the simulation data more similar to real data, and then use the transformed simulation data and measured data simultaneously for training the network.
(3) And (5) characteristic design.
A new hierarchical neural network was proposed by Dong et al in 2022, paper "Ahierarchical receptive network oriented to target recognition in SAR images" (Pattern recording). A signal-oriented receiver module is first built up by a series of fine convolution filters and then used to encode empirical features and knowledge. The final further refined features are used for model training.
(4) A domain adaptive method.
The paper "SAR target recognition based on task-driven domain adaptation using simulated data" (IEEE geoci. Remote. Sens. Lett) published by He et al in 2021 proposes a task-driven domain adaptation (TDDA) transfer learning method that can reduce the reduction in recognition capability caused by the difference in pitch angle between training and test data, by introducing relevant information of SAR imaging conditions at the time of training, for reducing the domain distribution difference between training and test data.
(5) Other methods.
Such as neural structure search, or explore the effectiveness of existing methods, and integrate some effective methods. Methods of data enhancement, model construction, loss function selection and combined use are proposed to enhance the feature information learned from simulation data by using an ATR model based on deep learning as in the paper "Bridging a gap in SAR-ATR: training on fully synthetic and testing on measured data" published by Inkawhich et al in 2021.
The above methods use more or less measured data containing labels for training in the training process, but it is still not easy to obtain enough measured SAR images for training at present, even for some complex tasks, even a small amount of measured data is difficult to obtain.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a progressive domain adaptive identification method for zero-sample SAR target identification. The technical problems to be solved by the invention are realized by the following technical scheme:
a progressive domain adaptive identification method for zero sample SAR target identification, the progressive domain adaptive identification method comprising:
step 1, acquiring a simulation data sample with a label and a measured data sample without a label;
step 2, respectively importing the labeled simulation data sample and the unlabeled measured data sample into a feature alignment network, and correspondingly extracting first feature information and second feature information;
step 3, calculating to obtain a distribution difference according to the first characteristic information and the second characteristic information, and taking the distribution difference as domain self-adaptive loss;
step 4, obtaining a first loss function according to the domain self-adaptive loss and a first classification loss corresponding to the labeled simulation data sample, so as to obtain a trained feature alignment network based on the first loss function;
step 5, predicting a label-free measured data sample by using the trained characteristic alignment network, and endowing a pseudo label to the label-free measured data sample to obtain a first measured data set of the pseudo label;
step 6, performing pseudo-tag denoising processing on the first measured data set to obtain a pseudo-tag denoising measured data set;
step 7, importing the actual measurement data set with the noise removed by the pseudo tag into the trained characteristic alignment network to adjust network parameters, and obtaining the adjusted characteristic alignment network;
and 8, inputting the data to be classified into the adjusted characteristic alignment network to obtain a final classification and identification result.
In one embodiment of the present invention, the step 1 includes:
step 1.1, adding speckle noise to a first simulation image to obtain a second simulation image;
step 1.2, cutting the first simulation image, the second simulation image and the first actually measured image to a uniform size, and correspondingly obtaining a third simulation image, a fourth simulation image and the second actually measured image;
and 1.3, normalizing the third simulation image, the fourth simulation image and the second actual measurement image to the range of [0,255] to obtain a fifth simulation image, a sixth simulation image and a third actual measurement image, wherein all the fifth simulation image with labels and the sixth simulation image with labels form the simulation data sample, and all the third actual measurement image without labels form the actual measurement data sample.
In one embodiment of the present invention, the feature alignment network includes a ten-layer network architecture, where a first layer is a convolution layer L1, a second layer is a maxPooling layer L2, a third layer is a convolution layer L3, a fourth layer is a maxPooling layer L4, a fifth layer is a convolution layer L5, a sixth layer is a maxPooling layer L6, a seventh layer is a convolution layer L7, an eighth layer is an avgPooling layer L8, a ninth layer is a convolution layer L9, and a tenth layer is a softmax classifier layer L10;
the convolution kernel sizes of the convolution layers L1, L3, L5 and L7 are 5×5, a dropout operation is added after the convolution layer L7, the convolution kernel size of the convolution layer L9 is 4×4, and finally, the identification result is output through the softmax classifier layer L10.
In one embodiment of the present invention, the first feature information and the second feature information are both feature information output by the convolution layer L7 of the feature alignment network.
In one embodiment of the invention, the domain adaptation loss is:
wherein ,D(XL ,X U ) For the domain adaptive loss, f (x l ) For the first characteristic information, f (x u ) E is mean value calculation for the second characteristic information;
the first loss function is:
L 1 ＝L C (X L )+λD(X L ,X U )
wherein ,L1 As a first loss function, L C (X L ) For the first classification loss, λ is the hyper-parameter.
In one embodiment of the present invention, the step 6 includes:
step 6.1, judging the relation between the probability of the predicted label output after the input features of the actual measurement data set of the pseudo label are aligned with the network and the threshold value, if the probability of the predicted label is smaller than or equal to the threshold value, removing the actual measurement data of the pseudo label, and if the probability of the predicted label is larger than the threshold value, reserving the actual measurement data of the pseudo label, wherein the reserved actual measurement data of all the pseudo labels form a second actual measurement data set;
step 6.2, establishing a similarity matrix based on the similarity of every two measured data in the second measured data set, and for each measured data in the second measured data set, finding a pseudo tag of the measured data with the highest similarity with the current measured data in the similarity matrix and marking as y tmax If the pseudo tag y of the current measured data t And y is tmax And if the current measured data is the same, adding the current measured data into the pseudo-tag denoising measured data set, otherwise, not adding the pseudo-tag denoising measured data set.
In one embodiment of the present invention, the similarity of the measured data includes cosine similarity, normalized mutual information similarity, or structural similarity.
In one embodiment of the present invention, the step 7 includes:
step 7.1, calculating to obtain second classification loss according to the actual measurement data set of the pseudo tag denoising;
and 7.2, obtaining a total loss function according to the second classified loss and the first loss function, so as to adjust network parameters of the trained feature alignment network based on the total loss function, and obtain an adjusted feature alignment network.
In one embodiment of the present invention, the second classification loss is:
wherein ,
pseudo tag for preliminary prediction ++>
A new prediction label after the actual measurement data set denoised by the pseudo label is adjusted in the characteristic alignment network;
the total loss function is:
L 2 ＝L C (X L )+λD(X L ,X U )+L C (X deno )
wherein ,L2 As a total loss function.
The invention has the beneficial effects that:
according to the invention, only the simulation data with the tag is used for identifying the actual measurement data, and the feature distribution difference between the simulation data and the actual measurement data is effectively reduced through the feature alignment network and the self-training method with pseudo tag denoising, so that the simulation data can be used for model training, the difficulty of acquiring the training data with the tag is reduced, and the time and economic cost for acquiring the actual measurement data are saved.
According to the progressive domain self-adaptive target recognition method for zero-sample SAR target recognition, a relatively stable primary recognition result can be obtained through the characteristic alignment network, and the self-training method with pseudo tag denoising effectively removes pseudo tag data containing noise, so that the recognition result is more stable.
Drawings
FIG. 1 is a schematic flow chart of a progressive domain adaptive recognition method for zero sample SAR target recognition according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a progressive domain adaptive network architecture provided by the present invention;
FIG. 3 is a diagram of a feature alignment network according to an embodiment of the present invention;
FIG. 4 is a process diagram of a method for implementing self-training according to an embodiment of the present invention;
fig. 5 is an explanatory diagram of a pseudo tag denoising method according to an embodiment of the present invention;
fig. 6 is a SAMPLE graph of a SAMPLE dataset for an experiment provided in an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to specific examples, but embodiments of the present invention are not limited thereto.
Example 1
Referring to fig. 1 and fig. 2, fig. 1 is a schematic flow chart of a progressive domain adaptive recognition method for zero-sample SAR target recognition according to an embodiment of the present invention; fig. 2 is a schematic diagram of a progressive domain adaptive network architecture provided by the present invention. The invention provides a progressive domain self-adaptive identification method for zero-sample SAR target identification, which comprises the following steps of 1-8, wherein:
step 1, acquiring a simulation data sample with a tag and a measured data sample without the tag, wherein the tag is used for identifying category information.
Step 1.1, adding speckle noise to a first simulation image to obtain a second simulation image;
step 1.2, clipping the first simulation image, the second simulation image and the first actually measured image to a uniform size, correspondingly obtaining a third simulation image, a fourth simulation image and the second actually measured image, for example clipping to a size of 64 multiplied by 64, so as to eliminate the background influence;
step 1.3, normalizing the third simulation image, the fourth simulation image and the second measured image to [0,255]]Obtaining a fifth simulation image, a sixth simulation image and a third actual measurement image in a range, wherein all the fifth simulation image with labels and the sixth simulation image with labels form a simulation data sample, all the third actual measurement image without labels form an actual measurement data sample, and the simulation data sample is marked as X L The measured data sample is marked as X U 。
And step 2, respectively importing the labeled simulation data sample and the unlabeled actual measurement data sample into a feature alignment network, and correspondingly extracting first feature information and second feature information.
In this embodiment, referring to fig. 3 for the structure of the feature alignment network Ω, the feature alignment network includes a ten-layer network architecture, where a first layer is a convolution layer L1, a second layer is a maxPooling layer L2, a third layer is a convolution layer L3, a fourth layer is a maxPooling layer L4, a fifth layer is a convolution layer L5, a sixth layer is a maxPooling layer L6, a seventh layer is a convolution layer L7, an eighth layer is an avgPooling layer L8, a ninth layer is a convolution layer L9, and a tenth layer is a softmax classifier layer L10;
the convolution kernel sizes of the convolution layers L1, L3, L5 and L7 are 5×5, a dropout (random inactivation) operation is added after the convolution layer L7, the convolution kernel size of the convolution layer L9 is 4×4, and finally the classification result is output through the softmax classifier layer L10.
Further, the first feature information and the second feature information are feature information output by a convolution layer L7 of the feature alignment network, namely a labeled simulation data sample X L And a sample X of measured data without label U And respectively extracting corresponding characteristic information of the two data samples after the two data samples pass through the convolution layer L7 through the characteristic alignment network.
And step 3, calculating to obtain distribution difference according to the first characteristic information and the second characteristic information, and taking the distribution difference as domain self-adaptive loss.
Specifically, the distribution difference is calculated by using the first characteristic information and the second characteristic information and is recorded as a domain adaptive loss D (X L ,X U ) The calculation formula is as follows:
wherein ,XL Representing a labeled dummy data sample, X U Represents a sample of measured data without labels, f (x l ) For the first characteristic information, f (x u ) And E is the average value for the second characteristic information.
And 4, obtaining a first loss function according to the domain self-adaptive loss and the first classification loss corresponding to the labeled simulation data sample, so as to obtain a trained feature alignment network based on the first loss function, namely obtaining the trained feature alignment network by minimizing the loss function.
Specifically, after training a simulated data sample containing a label through a feature alignment network, obtaining a corresponding classification loss L C (X L ) Then, the domain adaptive loss D (X) between the first characteristic information and the second characteristic information obtained in step 3 L ,X U ) A new loss function, i.e. a first loss function, is constructed, the calculation formula of which is as follows:
L 1 ＝L C (X L )+λD(X L ,X U )
wherein ,L1 As a first loss function, L C (X L ) For the first class of loss, λ is the superparameter, determining the domain adaptive loss to account for the loss L 1 Is a ratio of (2).
And 5, predicting a label-free measured data sample by using the trained characteristic alignment network, and endowing a pseudo label to the label-free measured data set to obtain a first measured data set of the pseudo label.
Specifically, after the trained features of the untagged actual measurement data sample are aligned to the network, a corresponding prediction tag is obtained, and the prediction tag is used as a pseudo tag, so that the actual measurement data set with the pseudo tag is used as a first actual measurement data set of the pseudo tag.
And 6, performing pseudo-tag denoising processing on the first measured data set to obtain a pseudo-tag denoising measured data set.
Because the pseudo tag inevitably has noise, i.e., the predicted tag may be erroneous, in order for the pseudo tag to be closer to the real tag, the present embodiment performs pseudo tag denoising processing on the first measured data set, specifically, may perform pseudo tag denoising processing through step 6.1 and step 6.2.
And 6.1, judging the relation between the probability of the predicted label output after the actual measurement data set input feature of the pseudo label is aligned with the network (namely, the result output by the softmax classifier layer L10 after the actual measurement data set input feature is aligned with the network) and the threshold, if the probability of the predicted label is smaller than or equal to the threshold, removing the actual measurement data of the pseudo label of which the probability of the predicted label is smaller than or equal to the threshold in the actual measurement data set, if the probability of the predicted label is larger than the threshold, retaining the actual measurement data of the pseudo label, and forming a second actual measurement data set by the retained actual measurement data of all the pseudo labels.
Specifically, a threshold value γ is set, i.e., measured data for a pseudo tag
If it predicts the probability P of the tag pred >And gamma, reserving the measured data of the pseudo tag for subsequent network fine adjustment, otherwise, not using the data for network fine adjustment.
Alternatively, γ is 0.9.
Step 6.2, establishing a similarity matrix based on the similarity of every two measured data in the second measured data set, finding a pseudo tag of the measured data (the measured data cannot be the current measured data) with the highest similarity with the current measured data in the similarity matrix for each measured data in the second measured data set, and marking as y tmax If the pseudo tag y of the current measured data t And y is tmax If the current measured data is the same, adding the current measured data into the pseudo tagAnd (5) denoising the actual measurement data set, otherwise, not adding.
Specifically, actually measured data of the pseudo tag selected by the threshold value gamma is not completely reliable, and in the embodiment, the pseudo tag is selected by combining image information besides setting the threshold value gamma.
Denoising process referring to fig. 4, the left plot in fig. 4 shows that there are some samples in the initial second measured data set that are mispredicted, and that the similarity of the samples in the same circle is relatively high, as indicated by the similarity measure. As can be seen from fig. 4, the predicted pseudo tag is guaranteed to be consistent between the two samples with the highest similarity. If the label is consistent, the pseudo label sample is reserved, otherwise, the pseudo label sample is rejected. Thus, based on the comparison of the similarity and its predicted pseudo tags, erroneous prediction samples can be eliminated, as shown in the right-hand diagram of FIG. 4. The pseudo tag denoising method specifically comprises the following steps:
establishing a similarity matrix S through all measured data in the second measured data set, wherein the similarity matrix S is as follows:
wherein n represents the number of measured data in the second measured data set, s tt ＝0，1≤t≤n。
For the measured data x t The pseudo tag is y t . For the measured data x t Finding the pseudo tag of the measured data with highest similarity in the similarity matrix S, and expressing the pseudo tag as y according to S tmax . Then judge the false label y tmax and yt If so, the measured data x is obtained t And adding the de-noised actual measurement data set, otherwise, not adding, and carrying out the processing on all the actual measurement pseudo tag data to obtain the final de-noised actual measurement data set.
In this embodiment, three different ways of cosine similarity (COSS), normalized Mutual Information Similarity (NMIS) and Structural Similarity (SSIM) are used to calculate the x-ray images 1 and x2 Between which are locatedIs briefly described as follows:
1)COSS:
wherein ,aij and bij respectively representing the images x 1 and x2 The pixel value at coordinates (i, j), n is the number of rows of the image and m is the number of columns of the image. Cosine similarity measures whether two images point to the same direction by calculating the cosine value of the included angle of the vectorized two image columns.
2)NMIS:
Wherein H (·) and H (·, ·) represent information and joint information entropy, respectively. Normalized mutual information similarity measures two images x from an entropy perspective 1 and x2 Similarity between them.
3)SSIM:
wherein ,μ1 Or mu 2 ，δ 1 Or delta 2 and δ12 Representing the mean, variance and covariance of the two images, respectively. Structural similarity measures the correlation between two images from the three parts of image brightness, contrast and structure. C (C) 1 、C 2 and C3 Is a constant set according to the related literature. C (C) 1 Is a constant for preventing zero removal by the image brightness comparing section, C 2 Is a constant for preventing zero division of the image contrast comparing section, C 3 Is a constant for preventing zero removal in the image structure comparison part, and is commonly taken as C 1 ＝0.0001，C 2 ＝0.0009，C 3 ＝0.5*C 2 。
And 7, importing the actual measurement data set denoised by the pseudo tag into a trained feature alignment network to adjust network parameters, and obtaining the adjusted feature alignment network.
That is, the present embodiment inputs the pseudo tag denoised measured data set to the trained feature alignment network to fine tune the network parameters.
Step 7.1, calculating a second classification loss L according to the measured data set of the pseudo tag denoising C (X deno ) Expressed as:
wherein ,Xdeno For the de-noised measured data set of the pseudo tag,
pseudo tag for preliminary prediction ++>
New predictive label for denoised pseudo-label measured data set after adjustment in feature alignment network,/for the pseudo-label denoised measured data set>
and />
Will be updated continuously with the network fine-tuning.
And 7.2, obtaining a total loss function according to the second classified loss and the first loss function, and adjusting the network parameters of the trained feature alignment network based on the total loss function to obtain an adjusted feature alignment network, namely fine-tuning the network parameters of the feature alignment network until the total loss function is minimum.
Referring to FIG. 5, in the trimming process, the total loss function L 2 Can be expressed as:
L 2 ＝L C (X L )+λD(X L ,X U )+L C (X deno )
wherein ,L2 As a total loss function.
And 8, inputting the data to be classified into the adjusted characteristic alignment network to obtain a final classification and identification result.
The invention aims to provide a progressive domain self-adaptive recognition framework for zero sample SAR target recognition aiming at the problem of distribution difference between a simulation image and an actual measurement image, and the progressive domain self-adaptive recognition framework comprises a feature alignment network and a self-training method with pseudo tag denoising so as to reduce the feature distribution difference between simulation data and actual measurement data and obtain better recognition performance.
The invention provides a progressive field self-adaptive target recognition framework for zero sample SAR target recognition, which reduces the feature distance between training data and test data through a training feature alignment network, predicts a label-free sample to obtain a pseudo label sample, and fine-adjusts the feature alignment classification network through a self-training method with pseudo label denoising to further improve the accuracy of recognition results.
According to the invention, only the simulation data with the tag is used for identifying the actual measurement data, and the feature distribution difference between the simulation data and the actual measurement data is effectively reduced through the feature alignment network and the self-training method with pseudo tag denoising, so that the simulation data can be used for model training, the difficulty of acquiring the training data with the tag is reduced, and the time and economic cost for acquiring the actual measurement data are saved.
According to the progressive field self-adaptive target recognition method for zero-sample SAR target recognition, a relatively stable primary recognition result can be obtained through the characteristic alignment network, and noise-containing pseudo tag data is effectively removed through the self-training method with pseudo tag denoising, so that the recognition result is more stable.
The effect of the invention can be further illustrated by the following experimental data:
1. experimental data
The dataset used in this experiment was a SAMPLE dataset containing a total of 10 types of data, each containing measured data and simulated data. The original image size was 128×128, and the resolution was 0.3m×0.3m. The pitch angle between the simulation data and the measured data is between 14 degrees and 17 degrees, wherein the simulation samples and the measured samples are in one-to-one correspondence. Table 1 presents the data types and data numbers of the original SAMPLE data set. This experimental scenario set-up is referred to as "training scenario 1". Referring to fig. 6, fig. 6 shows a sample of a pair of measured data and simulation data for each type. The lack of background clutter in the simulated image can be clearly seen. In addition, the strong scattering centers of the simulation data and the measured data are also different.
Table 1 training scenario 1: data type and number of SAMPLEs in original SAMPLE dataset
In the experiment of the invention, simulation data with the pitch angle ranging from 14 degrees to 16 degrees are used as training data, and measured data with the pitch angle ranging from 17 degrees are used as test data. This experimental setup was called "training scenario 2", and the pitch angle of the simulation data was different from that of the measured data, compared to "training scenario 1". Table 2 presents the specific data types and amounts of samples in "training scenario 2" for training and testing.
Table 2 sample data types and amounts for training and testing in training scenario 2
The experiment was trained using simulation data, measured data was tested. In the experiment, the data was cropped to a size of 64×64 at the time of image preprocessing to exclude background influence, and then normalized to the range of [0,255] by the min-max scaling method.
For network superparameters, the experiment uses a combination of parameters based on experimental exploration with a batch size of 256, a dropout of 0.2, a lambda of 10, and a total epoThe ch number was set to 500, and the learning rate was set to 10 using Adam optimizer -4 。
For the experimental results given, each set of experiments was performed 20 times, giving the lowest and highest recognition accuracy, denoted by "Min" and "Max", respectively. Ave represents the average recognition accuracy of 20 experiments and gives the standard deviation of the experimental results.
2. Evaluation criterion
The experimental results were evaluated using the following criteria:
recognition Accuracy Accuracy, minimum recognition Accuracy Min, maximum recognition Accuracy Max and average recognition Accuracy Ave.
The evaluation criteria described above are calculated using the following formula:
identification Accuracy Accuracy:
minimum recognition accuracy Min:
Min＝min(Accuracy 1 ,Accuracy 2 ,...,Accuracy N )
wherein n represents the number of times the same experiment is repeated under the same parameter setting, accuracy n The recognition accuracy of the nth experiment is shown, and min (·) represents the minimum value among given parameters.
Highest recognition accuracy Max:
Max＝max(Accuracy 1 ,Accuracy 2 ,...,Accuracy N )
wherein n represents the number of times the same experiment is repeated under the same parameter setting, accuracy n The recognition accuracy of the nth experiment is represented, and max (. Cndot.) represents the maximum value among given parameters.
Average recognition accuracy Ave:
in which n is as followsShows the number of times the same experiment was repeated under the same parameter setting, accuracy n The recognition accuracy of the nth experiment is shown.
3. Experimental details
Experiment one: the progressive domain adaptive framework provided by the invention is used for experiments, and as the framework is a two-stage training process, the characteristic alignment and self-training processes are required to be respectively carried out so as to verify the effectiveness of each stage. In the experiment, false label denoising is not performed. Tables 3 and 4 show experimental results, "baseline" is the way in which speckle noise is added to the image enhancement augmentation of the training data.
Table 3 training scenario 1: recognition results of progressive domain adaptation framework (%)
Table 4 training scenario 2: recognition results of progressive domain adaptation framework (%)
From the above results, it can be seen that the progressive domain adaptive framework proposed in the present invention works effectively in both training scenario settings, and the recognition results show a trend of stepwise improvement. In the second experiment, pseudo tag denoising is considered in the self-training process so as to obtain a better recognition result.
Experiment II: in this experiment, the denoising of the pseudo tag dataset was performed during the self-training process. And calculating the similarity between different images by using three similarity measures of COSS, NMIS and SSIM for pseudo tag noise reduction. Tables 5 and 6 show experimental results of a progressive domain adaptation framework for pseudo tag denoising.
Table 5 training scenario 1: identification result (%)
Table 6 training scenario 2: identification result (%)
And the false label denoising is carried out by adopting different similarity indexes, the average recognition performance can be improved by 1 to 2 percent compared with that before denoising, and the standard deviation of the multiple experiment results is reduced. In training scenario 1 and scenario 2, SSIM performs better and the average recognition rate is higher.
Experiment III: the framework of the present invention was compared to other existing frameworks using the SAMPLE dataset. Tables 7 and 8 show the experimental results of two training scenarios under different frame methods, respectively.
TABLE 7 training scenario 1 comparison with other methods (%)
TABLE 8 training scenario 2 comparison with other methods (%)
In contrast, the method in the invention achieves higher accuracy in the recognition result under the training scene 1, and in the training scene 2, the method in the invention also exceeds the results of other frames in the comparison experiment, and other researches utilize the SAMPLE data set to carry out experiments under different experimental settings, such as the experiment of Sellers et al in the paper "Augmenting simulations for SAR ATR neural network training" published in 2020, each class is trained by using one or two actually measured SAMPLEs besides the marked simulation data, and the method proposed in the experiment can achieve the classification accuracy of 95.1%, while the experimental result in the invention is still superior to the classification accuracy. Thus, the method of the present invention can be considered to have reached a more advanced level.
In summary, under the condition that only simulation data is used for training, a progressive domain self-adaptive framework is designed by adopting a pseudo-tag denoising method, experimental verification is carried out on a SAMPLE data set, and the framework effectively reduces the distribution difference between the simulation data and actual measurement data through a feature alignment network and a self-training mode containing pseudo-tag denoising, and obtains experimental results superior to other existing frameworks.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Further, one skilled in the art can engage and combine the different embodiments or examples described in this specification.
Although the present application has been described herein in connection with various embodiments, other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a review of the figures, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.
The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.
Claims (9)
1. A progressive domain adaptive identification method for zero sample SAR target identification, the progressive domain adaptive identification method comprising:
step 1, acquiring a simulation data sample with a label and a measured data sample without a label;
step 2, respectively importing the labeled simulation data sample and the unlabeled measured data sample into a feature alignment network, and correspondingly extracting first feature information and second feature information;
step 3, calculating to obtain a distribution difference according to the first characteristic information and the second characteristic information, and taking the distribution difference as domain self-adaptive loss;
step 4, obtaining a first loss function according to the domain self-adaptive loss and a first classification loss corresponding to the labeled simulation data sample, so as to obtain a trained feature alignment network based on the first loss function;
step 5, predicting a label-free measured data sample by using the trained characteristic alignment network, and endowing a pseudo label to the label-free measured data sample to obtain a first measured data set of the pseudo label;
step 6, performing pseudo-tag denoising processing on the first measured data set to obtain a pseudo-tag denoising measured data set;
step 7, importing the actual measurement data set with the noise removed by the pseudo tag into the trained characteristic alignment network to adjust network parameters, and obtaining the adjusted characteristic alignment network;
and 8, inputting the data to be classified into the adjusted characteristic alignment network to obtain a final classification and identification result.
2. The progressive domain adaptive identification method according to claim 1, wherein the step 1 comprises:
step 1.1, adding speckle noise to a first simulation image to obtain a second simulation image;
step 1.2, cutting the first simulation image, the second simulation image and the first actually measured image to a uniform size, and correspondingly obtaining a third simulation image, a fourth simulation image and the second actually measured image;
and 1.3, normalizing the third simulation image, the fourth simulation image and the second actual measurement image to the range of [0,255] to obtain a fifth simulation image, a sixth simulation image and a third actual measurement image, wherein all the fifth simulation image with labels and the sixth simulation image with labels form the simulation data sample, and all the third actual measurement image without labels form the actual measurement data sample.
3. The progressive domain adaptive identification method of claim 1, wherein the feature alignment network comprises a ten-layer network architecture, the first layer is a convolution layer L1, the second layer is a maxPooling layer L2, the third layer is a convolution layer L3, the fourth layer is a maxPooling layer L4, the fifth layer is a convolution layer L5, the sixth layer is a maxPooling layer L6, the seventh layer is a convolution layer L7, the eighth layer is an avgPooling layer L8, the ninth layer is a convolution layer L9, and the tenth layer is a softmax classifier layer L10;
the convolution kernel sizes of the convolution layers L1, L3, L5 and L7 are 5×5, a dropout operation is added after the convolution layer L7, the convolution kernel size of the convolution layer L9 is 4×4, and finally, the identification result is output through the softmax classifier layer L10.
4. A progressive domain adaptive identification method as defined in claim 3, wherein the first characteristic information and the second characteristic information are characteristic information output by the convolution layer L7 of the characteristic alignment network.
5. The progressive domain adaptive identification method of claim 1, wherein the domain adaptive loss is:
wherein ,D(XL ,X U ) For the domain adaptive loss, f (x l ) For the first characteristic information, f (x u ) E is mean value calculation for the second characteristic information;
the first loss function is:
L 1 ＝L C (X L )+λD(X L ,X U )
wherein ,L1 As a first loss function, L C (X L ) For the first classification loss, λ is the hyper-parameter.
6. The progressive domain adaptive identification method according to claim 1, wherein the step 6 comprises:
step 6.1, judging the relation between the probability of the predicted label output after the input features of the actual measurement data set of the pseudo label are aligned with the network and the threshold value, if the probability of the predicted label is smaller than or equal to the threshold value, removing the actual measurement data of the pseudo label, and if the probability of the predicted label is larger than the threshold value, reserving the actual measurement data of the pseudo label, wherein the reserved actual measurement data of all the pseudo labels form a second actual measurement data set;
step 6.2, establishing a similarity matrix based on the similarity of every two measured data in the second measured data set, and for each measured data in the second measured data set, finding a pseudo tag of the measured data with the highest similarity with the current measured data in the similarity matrix and marking as y tmax If the pseudo tag y of the current measured data t And y is tmax And if the current measured data is the same, adding the current measured data into the pseudo-tag denoising measured data set, otherwise, not adding the pseudo-tag denoising measured data set.
7. The progressive domain adaptive identification method of claim 6 wherein the similarity of the measured data comprises cosine similarity, normalized mutual information similarity, or structural similarity.
8. The progressive domain adaptive identification method of claim 5, wherein the step 7 comprises:
step 7.1, calculating to obtain second classification loss according to the actual measurement data set of the pseudo tag denoising;
and 7.2, obtaining a total loss function according to the second classified loss and the first loss function, so as to adjust network parameters of the trained feature alignment network based on the total loss function, and obtain an adjusted feature alignment network.
9. The progressive domain adaptive identification method of claim 7, wherein the second classification penalty is:
wherein ,
pseudo tag for preliminary prediction ++>
A new prediction label after the actual measurement data set denoised by the pseudo label is adjusted in the characteristic alignment network;
the total loss function is:
L 2 ＝L C (X L )+λD(X L ,X U )+L C (X deno )
wherein ,L2 As a total loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211711796.2A CN116030300A (en) | 2022-12-29 | 2022-12-29 | Progressive domain self-adaptive recognition method for zero-sample SAR target recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211711796.2A CN116030300A (en) | 2022-12-29 | 2022-12-29 | Progressive domain self-adaptive recognition method for zero-sample SAR target recognition |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116030300A true CN116030300A (en) | 2023-04-28 |
Family
ID=86080768
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211711796.2A Pending CN116030300A (en) | 2022-12-29 | 2022-12-29 | Progressive domain self-adaptive recognition method for zero-sample SAR target recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116030300A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117911852A (en) * | 2024-03-20 | 2024-04-19 | 西北工业大学 | Underwater target distance prediction method based on self-adaption in part of unsupervised field |
-
2022
- 2022-12-29 CN CN202211711796.2A patent/CN116030300A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117911852A (en) * | 2024-03-20 | 2024-04-19 | 西北工业大学 | Underwater target distance prediction method based on self-adaption in part of unsupervised field |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110472627B (en) | End-to-end SAR image recognition method, device and storage medium | |
| CN107563372B (en) | License plate positioning method based on deep learning SSD frame | |
| CN109558823B (en) | Vehicle identification method and system for searching images by images | |
| CN109215034B (en) | Weak supervision image semantic segmentation method based on spatial pyramid covering pooling | |
| CN107563433B (en) | Infrared small target detection method based on convolutional neural network | |
| CN112395987B (en) | SAR image target detection method based on unsupervised domain adaptive CNN | |
| CN106886760B (en) | A kind of EO-1 hyperion Ship Detection combined based on empty spectrum information | |
| CN111967480A (en) | Multi-scale self-attention target detection method based on weight sharing | |
| CN113392931B (en) | Hyperspectral open set classification method based on self-supervision learning and multitask learning | |
| CN106096506A (en) | Based on the SAR target identification method differentiating doubledictionary between subclass class | |
| CN103353988A (en) | Method for evaluating performance of heterogeneous SAR (synthetic aperture radar) image feature matching algorithm | |
| CN113011357A (en) | Depth fake face video positioning method based on space-time fusion | |
| CN110018453A (en) | Intelligent type recognition methods based on aircraft track feature | |
| CN108986142A (en) | Shelter target tracking based on the optimization of confidence map peak sidelobe ratio | |
| CN114926693A (en) | SAR image small sample identification method and device based on weighted distance | |
| CN110991257A (en) | Polarization SAR oil spill detection method based on feature fusion and SVM | |
| CN116030300A (en) | Progressive domain self-adaptive recognition method for zero-sample SAR target recognition | |
| Tang et al. | Target classification of ISAR images based on feature space optimisation of local non-negative matrix factorisation | |
| Sun et al. | A two-stage vehicle type recognition method combining the most effective Gabor features | |
| Huang et al. | EST-YOLOv5s: SAR Image Aircraft Target Detection Model Based on Improved YOLOv5s | |
| CN112613354A (en) | Heterogeneous remote sensing image change detection method based on sparse noise reduction self-encoder | |
| CN112270285A (en) | SAR image change detection method based on sparse representation and capsule network | |
| CN116704208A (en) | Local interpretable method based on characteristic relation | |
| CN112348750A (en) | SAR image change detection method based on threshold fusion and neighborhood voting | |
| CN111681266A (en) | Ship tracking method, system, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |