Semi-supervised small sample SAR image recognition system and method
Through the semi-supervised learning method combined with consistency regularity and pseudo-label cross-entropy minimization, the classification prediction and feature vector comparison of labelless data are solved, and the SAR image recognition algorithm performance in small sample scenarios is achieved, and a higher Identification accuracy and performance improvements.
Patent Information
- Application Number
- CN202410948203.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-16
Smart Images

Figure CN119027711A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of SAR image recognition, and in particular to a semi-supervised small sample SAR image recognition system and method. Background Art
[0002] Because it is not affected by climate and light, synthetic aperture radar remote sensing technology has important applications in battlefield reconnaissance, terrain mapping, geological exploration, ocean observation, disaster relief, etc. SAR automatic target recognition (SAR-ATR) usually consists of three steps: target detection, target extraction and target recognition. It is an important part of SAR image interpretation. As the last important part of SAR-ATR, SAR image recognition has always been the focus of research.
[0003] Traditional SAR image recognition algorithms first use domain knowledge to extract image features, and then classify and recognize them through machine learning algorithms, such as support vector machine (SVM), adaptive boosting algorithm (Adaboost), and sparse representation classification (SRC). The above methods have made great progress in SAR image recognition research, but they are heavily dependent on expert domain knowledge and manually extracted features, and the algorithm has poor migration and expansion capabilities.
[0004] With the continuous deepening of deep learning research, especially the breakthrough progress of convolutional neural network (CNN) in artificial intelligence applications, more and more people are trying to study the application of CNN algorithm in SAR image recognition and have made considerable achievements. The image recognition algorithm based on CNN extracts image features through convolution operation and continuously adjusts the model parameters through the feedback of classification error. This method performs feature extraction and classification at the same time, which not only improves performance, but also reduces the tediousness and difficulty of manual feature extraction.
[0005] However, the premise of using deep learning to automatically extract high-quality features is to have a large amount of training data. Compared with optical images, SAR images are more difficult to obtain, and the annotation of SAR images requires the knowledge of domain experts, which makes it difficult to obtain a large number of labeled SAR images. Therefore, studying how to improve the performance of SAR image recognition algorithms in the absence of labeled data has become an urgent problem to be solved.
[0006] The Chinese invention patent "A small sample SAR target recognition method based on hybrid attention mechanism" (application number 202310772323.1) uses a multi-scale feature fusion strategy to fuse multiple layers of features, and introduces a hybrid attention mechanism into a densely connected network to improve the accuracy of small sample target recognition. The Chinese invention patent "A small sample SAR target recognition method based on twin neural network" (application number 202310942393.7) uses DenseNet and Inception technology in the twin network backbone network, which effectively improves the image recognition accuracy when the data set is small, the category is single, and the scene is complex. Although the above methods improve the performance of the model by designing the network structure of the SAR image recognition model, they only use labeled data. In reality, due to the difficulty of labeling, a large amount of data is unlabeled. The Chinese invention patent "SAR target recognition method based on self-supervised learning and sub-aperture decomposition" (application number 202310581453.7) expands the unlabeled data into two types of data including sub-aperture decomposition, inputs them into different networks, and trains them by comparing the similarity between the two. Although this method further utilizes unlabeled data, it only compares the feature outputs of different data expansion versions of the same sample, does not fully utilize the classification outputs of the two, and does not consider the mixed comparison of labeled and unlabeled sample data. Summary of the invention
[0007] The purpose of the present invention is to provide a semi-supervised small sample SAR image recognition system and method. The present invention combines consistency regularization with pseudo-label-based cross entropy minimization, simultaneously evaluates model loss from two aspects: data features and classification prediction, fuses labeled data and unlabeled data information, explores the correlation between unlabeled data and labeled data, and fully mines the information of unlabeled data to improve the performance of the model.
[0008] To achieve this purpose, the present invention designs a semi-supervised small sample SAR image recognition system, which includes a model training module and an image recognition module;
[0009] The model training module is used to perform data enhancement on the SAR image training data in batches, and transmit the data-enhanced SAR image training data to the SAR image recognition model composed of the feature extraction network and the classifier, and calculate the loss of the SAR image recognition model according to the loss function to complete the training of the SAR image recognition model;
[0010] The image recognition module is used to transmit the SAR image to be recognized to the trained SAR image recognition model for SAR image recognition.
[0011] Beneficial effects of the present invention:
[0012] The present invention designs a semi-supervised learning method. The algorithm combines consistency regularization and a cross entropy minimization method based on pseudo-labels. By comparing the classification predictions and feature vectors of high-confidence unlabeled samples with strong and weak data enhancement, the model loss is estimated from multiple aspects, the model's sensitivity to pseudo-label errors is reduced, and the error accumulation problem is alleviated. By designing a distance metric loss that combines labeled and unlabeled data, the model is optimized in the direction of intra-class aggregation and inter-class dispersion. In the distance metric method, by storing the global feature vector method in real time, a more accurate category prototype is obtained without increasing the amount of training calculations. The unsupervised algorithm designed by the present invention fully mines the information of unlabeled samples, greatly improves the accuracy of SAR image recognition in small sample scenarios, and alleviates the difficulty of SAR image recognition caused by insufficient samples. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a structural schematic diagram of the present invention;
[0014] Figure 2 It is the overall framework diagram of the algorithm of the present invention;
[0015] Figure 3 Schematic diagram of the prototype for merging labeled and unlabeled data calculations;
[0016] Figure 4 The output features of the models trained by different algorithms for the test data set are visualized after t-SNE dimensionality reduction: (a) is the algorithm in this paper; (b) is the fixmatch algorithm; (c) is the supervised training algorithm with only labeled data;
[0017] Figure 5 This is a line chart comparing the recognition rates of different algorithms. DETAILED DESCRIPTION
[0018] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0019] Example 1
[0020] A semi-supervised small sample SAR image recognition system, such as Figure 1 and 2 As shown, it includes a model training module and an image recognition module;
[0021] The model training module is used to perform data enhancement on the SAR image training data in batches (the number of samples in each batch is set equal), and transmit the data-enhanced SAR image training data to the SAR image recognition model composed of the feature extraction network and the classifier, and calculate the loss of the SAR image recognition model according to the loss function to complete the training of the SAR image recognition model;
[0022] The image recognition module is used to transmit the SAR image to be recognized to the trained SAR image recognition model for SAR image recognition.
[0023] In the above technical solution, the data enhancement method is: weak enhancement is performed on samples with image category labels in the SAR image training data, and weak enhancement and strong enhancement are performed on samples without image category labels, and the above weak enhancement and strong enhancement are performed in parallel.
[0024] In the above scheme, labeled means that the training samples are pictures and the category indicators corresponding to the pictures; unlabeled means that the training samples only have pictures, and there is no information about which category the pictures belong to. In classification problems, the meaning of the label is the category to which the data actually belongs.
[0025] In the above technical solution, the weak enhancement process is as follows: first, the samples in the SAR image training data are center-cropped (a region of a specified size is cropped outward from the center point of the image) into a uniform size (88*88), and then randomly flipped vertically and horizontally with a probability of 50%, and finally the pixel grayscale values are uniformly normalized to a mean of 0.485, and the standard deviation of the pixel value of the image is 0.229;
[0026] The strong enhancement process is as follows: first, the samples in the SAR image training data are center-cropped into a uniform size (88*88), and then randomly flipped vertically and horizontally with a probability of 50%. Then, three types are randomly selected from automatic contrast, color transformation, contrast (automatic contrast does not require setting parameters, automatically adjusts the contrast of the image to enhance the image contrast; contrast adjusts the image contrast according to given parameters), equivalence, identity, posterize, rotate, adjust sharpness, shear, solarize, and translate for random enhancement. Finally, the pixel grayscale values are uniformly normalized to a mean of 0.485 and a standard deviation of 0.229.
[0027] In the above technical solution, the data-enhanced SAR image training data is transmitted to the SAR image recognition model composed of a feature extraction network and a classifier:
[0028] The enhanced SAR image training data is input into the feature extraction network f φ (·), and obtain the feature vector f with weak enhancement of image category label samples i , the feature vector f of the weakly enhanced sample without image category label iw And the feature vector f strongly enhanced by samples without image category labels is , the feature vector f i , eigenvector fiw and the eigenvector f is Input classifier g ψ (·), and obtain the eigenvector f i , eigenvector f iw and the eigenvector f is The classification prediction value p i 、p iw and p is , using the classification prediction value p iw , use the softmax function to calculate the soft label of the unlabeled sample weak enhancement By calculating the peak position in the soft label, the pseudo label is obtained The pseudo label indicates which category has the largest prediction value, that is, which category has the highest probability of being assigned to it. It is not the real label, but the label calculated by the algorithm. The feature extraction network uses restnet_18, and the classifier uses a linear layer from 512 to 10.
[0029] Soft labels for weak enhancement of unlabeled samples The peak value is compared with the set threshold, and the samples exceeding the threshold are recorded as high-confidence samples without image category labels.
[0030] The feature vector f of all samples with image category labels in this batch is weakly enhanced i Store in the global storage M_fx. If the storage is full, replace the oldest batch of data. Store the high-confidence weakly enhanced feature vectors without image category labels in this batch in the global storage M_fu. If the storage is full, replace the oldest batch of data. Use the data in the global storage to calculate the feature prototype C (feature vector center) of each category. The global storage contains the feature vector data of the most recent batches. The feature prototype calculated using the data in the global storage is more consistent with the actual data distribution than the prototype calculated using only the current batch.
[0031] Using the eigenvector f i 、f iw and f is , classification prediction value p i 、p iw 、p is , pseudo label High confidence samples without image category labels The loss of the SAR image recognition model is calculated comprehensively with the feature prototype C.
[0032] In the above technical solution, high confidence samples without image category labels The specific calculation formula is:
[0033]
[0034] Among them, U b ={u1,...,u μB} is the set of all unlabeled samples in the current batch of SAR image training data, u μB represents the μBth sample without image category label, τ represents the high confidence threshold, u i Represents the i-th sample without image category label in this batch.
[0035] In the above scheme, the global storage space is the computer memory space allocated in the computer before model training and can be accessed for multiple rounds of training. The global storage M_fx is used to store the feature vectors of samples with image labels, and m1 slots are allocated, with each batch of data occupying 1 slot; the global storage M_fu is used to store the feature vectors of samples without image labels, and m2 slots are allocated, with each batch of data occupying 1 slot; where m1 and m2 are the number of batches required to traverse the labeled data and unlabeled data, respectively, and the specific calculation formula is:
[0036] m1=N x / B
[0037] m2=N u / (μB)
[0038] Among them, N x is the number of samples with image category labels, N u is the number of all samples without image category labels, B is the number of labeled samples in each batch of SAR image training data, and μ is the ratio of the number of unlabeled samples to labeled samples in each batch of SAR image training data.
[0039] In the above technical solution, the calculation process of the loss Loss of the SAR image recognition model in the loss function is calculated according to the following formula:
[0040] Loss = L x +αL u +βL f +γL dis
[0041] in, Lx and Lu They are respectively the data loss with image category labels and the sample loss with high confidence without image category labels. Ldis is the distance metric loss between samples with image category labels and samples without image category labels with high confidence in the current batch and their prototypes, L f is the characteristic loss of the SAR image recognition model, α, β and γ are the control Lu , Lf and Ldis The hyperparameters of the three loss ratios are set to 1 for α, 1.2 for β, and 2.5 for γ;
[0042] The loss of the SAR image recognition model is back-propagated, and the parameters of the SAR image recognition model are updated using the stochastic gradient descent algorithm SGD to complete the training of the SAR image recognition model consisting of the feature extraction network and the classifier. The learning rate in the stochastic gradient descent algorithm is 0.03, the momentum is 0.9, and the cosine scheduler is used to adjust the learning rate.
[0043] In the above technical solution, the image category label data is lost Lx The calculation method is:
[0044]
[0045] Among them, B is the number of samples with image category labels in each batch of SAR image training data, and p i is a classifier for samples x with image category labels i The classification output, y i For x i is the image category label, H(·) is the cross entropy function, and i is the sample number;
[0046] High confidence loss of samples without image class labels Lu The calculation method is:
[0047]
[0048] in, is a sample u without image category label i The pseudo label of the weakly enhanced classification output, p is is a sample u without image category label i The strongly enhanced classification output, is a high confidence sample without image category label, is the number of elements in the high confidence samples without image category labels;
[0049] Feature loss L of SAR image recognition model f The calculation method is:
[0050]
[0051] Let the feature vector set stored in the global storage M_fx be denoted by F x , let the feature vector set stored in the global storage M_fu be F x and The data is merged and recorded as set F. Set F is combined according to image categories. The set of all feature vectors of the lth category is recorded as F(l). The prototype c of the lth category is calculated according to the following formula l , which is the eigenvector center:
[0052]
[0053] Where, f is an element in F(l);
[0054] The metric loss L of the distance between the samples with image category labels and the samples without image category labels with high confidence in the previous batch and their prototypes dis The calculation method is:
[0055]
[0056] Among them, L dis-x Indicates the distance loss of image category label data, L dis-u represents the distance loss of data without image category labels, L represents the image category label set, d(·) is the distance measurement formula, and the present invention uses the Euclidean distance formula, d(f iw ,c l ) represents a sample u without image category label i After weak enhancement, the feature vector f is obtained by the feature extraction network iw and the prototype c of the category to which this sample belongs l The distance, d(f iw ,c l′ ) represents a sample u without image category label i After weak enhancement, the feature vector f is obtained by the feature extraction network iw and the prototype c of the l′th class l′ distance.
[0057] In the above scheme, the global storage space is the computer memory space allocated in the computer before model training and can be accessed for multiple rounds of training.
[0058] The above global storage is allocated as follows:
[0059] The global storage M_fx is used to store the feature vectors of samples with image labels, and m1 slots are allocated, with each batch of data occupying 1 slot. The global storage M_fu is used to store the feature vectors of samples without image labels, and m2 slots are allocated, with each batch of data occupying 1 slot. Among them, m1 and m2 are the number of batches required to traverse the labeled data and the unlabeled data, respectively. The specific calculation formula is:
[0060] m1=N x / B
[0061] m2=N u / (μB)
[0062] Where N x is the number of all labeled SAR image samples, N uis the number of all unlabeled samples, B is the number of labeled samples in each batch, and μ is the ratio of the number of unlabeled samples to labeled samples in each batch (the number of samples in each batch is set equal). Labeled means that the training samples are images and the corresponding category indicators of the images, and unlabeled means that the training samples only have images, and there is no information about which category the images belong to. In classification problems, the meaning of the label is the category to which the data actually belongs.
[0063] The global storage design idea is:
[0064] The present invention uses a distance measurement algorithm to measure the distribution of samples in feature space. The distance measurement algorithm calculates the model loss by measuring the distance between samples of the same type and samples of different types in the feature vector output by the SAR image recognition model, and is an effective method for improving the performance of the image recognition model. The present invention uses a classic prototype network distance measurement method, first calculates the prototype (center) of the feature vector of each image category, and then sets the distance (Euclidean distance) from each sample in each category to its prototype as the model loss, and continuously optimizes the parameters of the neural network so that the distance from each sample to its prototype becomes smaller and smaller, so that the model is optimized in the direction of intra-class aggregation and inter-class distance.
[0065] However, in the small sample scenario, the number of labeled samples is insufficient, making it difficult to form statistical characteristics, and the calculated category prototype is quite different from the actual category prototype. See the schematic diagram Figure 3 . The light blue solid triangles are a certain type of labeled data, which are usually less in number. The hollow triangles are unlabeled data, which are more in number. The dark blue triangle c is the prototype (center) of this type of labeled data, and the green triangle c′ is the prototype of this type of mixed labeled and unlabeled data. Since the amount of labeled data is too small, there is a certain deviation between its distribution (inner dotted circle range) and the overall data distribution (outer dotted circle range), so the category prototype calculated only with labeled data cannot represent the true prototype of this category. By merging labeled samples and high-confidence unlabeled samples (calculating the average value of labeled and unlabeled sample data together), the algorithm of the present invention can effectively increase the number of samples required for prototype calculation, and the calculated prototype is more accurate.
[0066] However, in the early stage of SAR image recognition model training, there are fewer high-confidence unlabeled samples in each batch. In order to make the prototype calculation more accurate, the samples involved in the prototype calculation can be not limited to the current batch, but can use multiple batches or even all available data. However, calculating the feature vectors of all samples in each batch will greatly increase the amount of calculation and time for model training. In order to simultaneously meet the requirements of utilizing more sample data without seriously increasing the amount of training calculation, the present invention introduces an online feature method. Unlike recalculating the feature vectors of all samples in each batch, the online feature method uses global storage to store the feature vectors of the most recent batches. By updating the data in the global storage in each batch, the global storage maintains the storage of the most recently output features, which can be approximately considered as the features calculated using the current model, thereby eliminating the need to recalculate the features of all data each time, greatly reducing the amount of calculation in the training process.
[0067] The introduction of multiple batches of feature vectors increases the data supporting prototype calculations, but what strategy is used to mix labeled and unlabeled data requires further study. In order to be closer to the real data distribution, the present invention does not simply merge each batch of labeled data and unlabeled data, but keeps their proportions consistent with the overall training data. The specific implementation strategy is as follows: suppose that all labeled and unlabeled training data require m1 and m2 batches to traverse respectively. In a small sample scenario, m2>m1, then the number of batches of labeled and unlabeled data stored in the global storage is the most recent m1 and m2 times, so the present invention uses 2 global storages, and the global storage M_fx stores the feature vectors of labeled sample data, and the slot size is m1 (each slot stores a batch of data). The global storage M_fu stores the feature vectors of high-confidence unlabeled sample data, and the slot size is m2 (each slot stores a batch of data). When calculating the distance metric, the loss ratio of labeled and unlabeled data is also m1:m2. The prototype and distance loss are calculated in the same proportion according to all training data distribution, instead of simply merging labeled data with unlabeled data in each iteration. This avoids the problem that the prototype deviates from the distribution of labeled data and differs greatly from the actual situation due to the reuse of labeled data.
[0068] The present invention trains the SAR image recognition model for all SAR image training data in batches and repeats this process 300 times to obtain a final SAR image recognition model, and then uses the final SAR image recognition model to perform SAR image recognition.
[0069] Example 2
[0070] A semi-supervised small sample SAR image recognition method, the process of the method is as follows: training stage: first, data enhancement is performed on labeled training data and unlabeled training data; then the enhanced data is input into a feature extraction network and a classifier; the model loss is obtained by calculating the error between labeled data and unlabeled data in feature vectors and classification predictions; and the model parameters are continuously optimized by minimizing the loss. Testing stage: the image to be recognized is input into the trained model to obtain the recognition result. In the loss function design of the present invention, the entropy minimization and consistency regularization methods based on pseudo labels are combined, the differences in classification prediction and feature vectors of unlabeled data of different enhancement types are compared, and a distance measurement algorithm combining labeled and unlabeled data is designed to fully mine the information contained in the unlabeled samples, and alleviate the problem of poor performance of deep learning models caused by insufficient labeled samples of SAR images. By comparing with supervised learning algorithms, other algorithms in small sample scenarios in the SAR-ATR field proposed in recent years, and other semi-supervised learning algorithms, the superiority of the method of the present invention is fully demonstrated.
[0071] The specific steps of the semi-supervised small sample SAR image recognition method are:
[0072] The SAR image training data is enhanced in batches, and the enhanced SAR image training data is transmitted to the SAR image recognition model composed of a feature extraction network and a classifier, and the loss of the SAR image recognition model is calculated according to the loss function to complete the training of the SAR image recognition model;
[0073] The SAR image to be identified is transmitted to the trained SAR image recognition model for SAR image recognition.
[0074] Example 3
[0075] A computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the method described in Example 2 are implemented.
[0076] The effect of the present invention can be further illustrated by the following experiment, and the experimental process is as follows:
[0077] Experimental conditions
[0078] The experiment selected 10 types of ground target data in the standard operating environment (SOC) of the MSTAR dataset for the experiment, and the distribution of experimental data is shown in Table 1. In order to fully test the effect of the algorithm in the small sample scenario, this paper randomly selected 5, 10, 20, and 30 samples from each category to form a labeled sample dataset, and the remaining training sets constituted an unlabeled sample dataset.
[0079] The experiment was carried out under Ubuntu 20.04 LTS, using the pytorch platform, the graphics card configuration was RTX 3090, and the experimental parameters were configured as follows: the feature extractor used Restnet-18, and the output dimension was 512; the classifier was a single-layer linear classifier, the input dimension was 512, and the output dimension was 10; the input image size was 88*88; the batch size was 64; the SGD optimizer was selected, the initial learning rate was 0.03, and the cosine annealing algorithm was used to adjust the learning rate. The loss function hyperparameters α, β, and γ were set to 1, 1.2, and 2.5, respectively.
[0080] Table 1 MSTAR SOC training set and test set
[0081]
[0082] Experimental content
[0083] Experiment 1
[0084] The training set and test set described in the experimental conditions are used to train and test the algorithm of the present invention and the supervised learning algorithm, and the supervised learning algorithm uses only labeled data. Table 2 is a comparison of the experimental results. It can be seen from the results that after using unlabeled data, the algorithm performance is greatly improved, especially for the case where the number of samples is particularly small, the algorithm improvement is more obvious. When there are only 30 data (about 10%) in each category, the accuracy rate can reach 99%. When the number of samples is extremely small (only 5 samples in each category), the recognition rate can reach more than 85%, which is 68% higher than the performance of using only labeled data. This proves that the algorithm in this paper makes full use of unlabeled data to improve model performance, which can greatly alleviate the problem of small labeled data and low image recognition rate in real situations.
[0085] Table 2 Comparison of recognition rate experimental results between the proposed algorithm and the algorithm using only labeled data
[0086]
[0087] Experiment 2
[0088] The training set and test set described in the experimental conditions are used to train and test the algorithm in this paper and the fixmatch (Sohn K, Berthelot D, Li CL, et al. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence [J]. 2020. DOI: 10.48550 / arXiv.2001.07685.) algorithm. The algorithm in this paper is improved on the basis of fixmatch. Table 3 compares the experimental results of the algorithm in this paper and the fixmatch algorithm under the same training conditions. It can be seen from the results that both perform well in small sample scenarios. In the case of very few samples, the algorithm in this paper shows obvious advantages. This is because in addition to using the supervised loss of labeled samples and the soft label of unlabeled samples proposed by the benchmark algorithm, this paper also uses the feature vector difference and the distance metric of labeled and unlabeled merge, comprehensively considering the model loss from multiple aspects, reducing the sensitivity of the model to erroneous soft labels, and improving the model performance.
[0089] Table 3 Comparison of recognition rate experimental results between this algorithm and fixmatch algorithm
[0090] 5 10 20 30 100 fixmatch 67.68 94.3 98.13 98.21 98.92 This article's algorithm 86.56 97.57 98.00 99.33 99.36
[0091] Figure 4 It is a visualization of the features output by the test data set under the models trained by different algorithms after t-SNE dimension reduction, where (a) is the algorithm of the present invention, (b) is the fixmatch algorithm, and (c) is a supervised training algorithm using only labeled data. It can be seen from the figure that the features output by the algorithm in this paper are obviously more clustered than other algorithms, and the classification effect is better. This is because the algorithm of the present invention not only uses a large amount of unlabeled data for semi-supervised learning, but also introduces a distance metric that combines labeled and unlabeled data, so that the model is optimized in the direction of intra-class aggregation and inter-class distance. The effectiveness of the distance measurement method proposed in this paper can also be proved from the feature visualization diagram.
[0092] Experiment 3
[0093] To further demonstrate the advantages of the algorithm of the present invention, we compared the algorithm proposed in other recent literature with the algorithm of the present invention. Algorithm for SAR image recognition in small sample scenarios:
[0094] Domain knowledge-powered two-stream deep network(DKTS-N)(Zhang,L.,Leng,X.,Feng,S.,Ma,X.,Ji,K.,Kuang,G.,and Liu,L.:Domain knowledge powered two-stream deep network for few-shot SAR vehicle recognition.IEEE Transactions onGeoscience and Remote Sensing.60,1-15(2021).);
[0095] Cascaded Features and AdaBoost Rotation Forest(CF_ARF)(Zhang,F.,Wang,Y.,Ni,J.,Zhou,Y.,and Hu,W.:SAR target small sample recognition based on CNNcascaded features and AdaBoost rotation forest.IEEE Geoscience and RemoteSensing Letters.17(6),1008-1012(2019).);
[0096] hierarchically designed lightweight method(HDLM)(Wang,C.,Pei,J.,Yang,J.,Liu,X.,Huang,Y.,and Mao,D.:Recognition in label and discrimination infeature:A hierarchically designed lightweight method for limited data in saratr.IEEE Transactions on Geoscience and Remote Sensing.60,1-13(2022).);
[0097] Deep Kernel Learning (DKL) (Wang, K., Qiao, Q., Zhang, G., and Xu, Y.: Few-shot SAR target recognition based on deep kernel learning. IEEE Access. 10, 89534-89544 (2022).);
[0098] There is the distance metric loss algorithm two-stage multiscale densely connected convolutional neural networks (TMDC-CNNS) (Guan, J., Liu, J., Feng, P., and Wang, W.: Multiscale deep neural network with two-stage loss for SAR target recognition with small training set. IEEE geoscience and remote sensing letters. 19, 1-5 (2021).);
[0099] Classification Using CNN Embeddings and Metric Learning (CEML) (Li, Y., Li, X., Sun, Q., and Dong, Q.: SAR image classification using CNN embeddings and metric learning. IEEE Geoscience and Remote Sensing Letters. 19, 1-5 (2020).);
[0100] The semi-supervised algorithm Self-Consistent Augmentation (SCA) (Wang, C., Shi, J., Zhou, Y., Yang, X., Zhou, Z., Wei, S., and Zhang, X.: Semisupervised learning-based SAR ATR via self-consistent augmentation. IEEE Transactions on Geoscience and Remote Sensing. 59(6), 4862-4873 (2020).);
[0101] Table 4 gives the comparison results. The results of the compared algorithms are all taken from the original literature. It can be seen that the proposed algorithm shows the best performance in a variety of small sample scenarios. When the number of samples is very small (5 or 10 samples per category), the proposed algorithm shows obvious advantages.
[0102] Table 4 Comparison of experimental results of this algorithm and other literature algorithms
[0103] 5 10 20 30 TMDC-CNNS 64.11 80.17 - 95.14 CEML - - 82.29 - DKTS-N - 87 93 97.1 CF_ARF - 77.5 85.9 92.2 (HDLM) 72.37 88.66 95.17 - SCA - 95.22 98.23 - DKL 69.14 79.45 - - fixmatch 67.68 94.3 98.13 98.21 This article's algorithm 86.56 97.57 98.00 99.33
[0104] Figure 5 It is a line graph comparing the recognition rates of various algorithms under different sample numbers. As can be seen from the figure, the algorithm in this paper shows obvious advantages when there are few samples, and has a higher improvement rate as the number of samples increases when there are fewer samples. The algorithm in this paper, fixmatch and SCA algorithms all use semi-supervised learning methods, which perform better than other algorithms in small sample scenarios, and achieve a recognition rate of more than 98% when there are only 20 samples (approximately) in each category. This also proves the effectiveness of semi-supervised learning algorithms for small sample scenarios. The algorithm in this paper has obvious advantages over other algorithms in the case of very few samples because the algorithm in this paper designs the loss function of feature vector comparison and distance measurement on the basis of the semi-supervised algorithm, which makes fuller use of unlabeled data information and reduces the error impact of unlabeled data.
[0105] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.
Claims
1. A semi-supervised small sample SAR image recognition system, Features: It includes a model training module and an image recognition module; The model training module is used to perform data enhancement on the SAR image training data in batches, and transmit the data-enhanced SAR image training data to the SAR image recognition model composed of the feature extraction network and the classifier, and calculate the loss of the SAR image recognition model according to the loss function to complete the training of the SAR image recognition model; The image recognition module is used to transmit the SAR image to be recognized to the trained SAR image recognition model for SAR image recognition.
2. The semi-supervised small sample SAR image recognition system according to claim 1, Features: The data enhancement method is as follows: weak enhancement is performed on samples with image category labels in the SAR image training data, and weak enhancement and strong enhancement are performed on samples without image category labels.
3. The semi-supervised small sample SAR image recognition system according to claim 2, Features: The weak enhancement process is as follows: firstly, samples in the SAR image training data are centrally cropped into a uniform size, then randomly flipped vertically and horizontally, and finally, the pixel grayscale values are uniformly normalized to a mean value; The strong enhancement process is as follows: first, the samples in the SAR image training data are center-cropped into a uniform size, and then randomly flipped vertically and horizontally. Then, three types of random enhancement are randomly selected from automatic contrast, color transformation, contrast, equivalence, identity transformation, tone separation, rotation, sharpness adjustment, shearing, exposure, and translation. Finally, the pixel grayscale values are uniformly normalized to the mean.
4. The semi-supervised small sample SAR image recognition system according to claim 1, Features: The data-enhanced SAR image training data is fed into the SAR image recognition model consisting of a feature extraction network and a classifier: The enhanced SAR image training data is input into the feature extraction network f φ (·), and obtain the feature vector f with weak enhancement of image category label samples i , the feature vector f of the weakly enhanced sample without image category label iw And the feature vector f strongly enhanced by samples without image category labels is , the feature vector f i , eigenvector f iw and the eigenvector f is Input classifier g ψ (·), and obtain the eigenvector f i , eigenvector f iw and the eigenvector f is The classification prediction value p i 、p iw and p is , using the classification prediction value p iw , use the softmax function to calculate the soft label of the unlabeled sample weak enhancement By calculating the peak position in the soft label, the pseudo label is obtained Soft labels for weak enhancement of unlabeled samples The peak value is compared with the set threshold, and the samples exceeding the threshold are recorded as high-confidence samples without image category labels. The feature vector f of all samples with image category labels in this batch is weakly enhanced i Store in the global storage M_fx. If the storage is full, replace the oldest batch of data. Store the high-confidence weakly enhanced feature vectors without image category labels in this batch in the global storage M_fu. If the storage is full, replace the oldest batch of data. Use the data in the global storage to calculate the feature prototype C of each category. The global storage contains the feature vector data of the most recent batches. Using the eigenvector f i 、f iw and f is , classification prediction value p i 、p iw 、p is , pseudo label High confidence samples without image category labels The loss of the SAR image recognition model is calculated comprehensively with the feature prototype C.
5. The semi-supervised small sample SAR image recognition system according to claim 4, Features: High confidence samples without image category labels The specific calculation formula is: Among them, U b = {u 1 ,...,u μB } is the set of all unlabeled samples in the current batch of SAR image training data, u μB represents the μBth sample without image category label, τ represents the high confidence threshold, u i Represents the i-th sample without image category label in this batch.
6. The semi-supervised small sample SAR image recognition system according to claim 4, Features: The global storage M_fx has a total of m 1 slots, the global storage M_fu has a total of m 2 slots, of which m 1 and m 2 are the number of batches required to traverse the labeled data and unlabeled data respectively; m 1 =N x / B m 2 =N u / (μB) Among them, N x is the number of samples with image category labels, N u is the number of all samples without image category labels, B is the number of labeled samples in each batch of SAR image training data, and μ is the ratio of the number of unlabeled samples to labeled samples in each batch of SAR image training data.
7. The semi-supervised small sample SAR image recognition system according to claim 4, Features: The calculation process of the loss Loss of the SAR image recognition model in the loss function is calculated according to the following formula: Loss=L x +αL u +βL f +γL dis in, Lx and Lu They are respectively the data loss with image category labels and the sample loss with high confidence without image category labels. Ldis is the metric loss of the distance between the samples with image category labels and the samples without image category labels with high confidence in the current batch and their prototypes, L f is the feature loss of the SAR image recognition model; The loss of the SAR image recognition model is back-propagated, and the parameters of the SAR image recognition model are updated using the stochastic gradient descent algorithm to complete the training of the SAR image recognition model composed of the feature extraction network and the classifier. α, β, and γ are the control Lu , Lf and Ldis Hyperparameters for the three loss ratios.
8. The semi-supervised small sample SAR image recognition system according to claim 7, Features: The image category label data loss Lx The calculation method is: Among them, B is the number of samples with image category labels in each batch of SAR image training data, and p i is a classifier for samples with image category labels x i The classification output, y i For x i is the image category label, H(·) is the cross entropy function, and i is the sample number; High confidence loss of samples without image class labels Lu The calculation method is: in, is a sample u without image category label i The pseudo label of the weakly enhanced classification output, p is is a sample u without image category label i The strongly enhanced classification output, is a high confidence sample without image category label, is the number of elements in the high confidence samples without image category labels; Feature loss L of SAR image recognition model f The calculation method is: Let the feature vector set stored in the global storage M_fx be denoted by F x , let the feature vector set stored in the global storage M_fu be F x and The data is merged and recorded as set F. Set F is combined according to image categories. The set of all feature vectors of the lth category is recorded as F(l). The prototype c of the lth category is calculated according to the following formula l : Where, f is an element in F(l); The metric loss of the distance between the samples with image category labels and the samples without image category labels with high confidence in the previous batch and their prototypes Ldis The calculation method is: Among them, L dis-x Indicates the distance loss of image category label data, L dis-u represents the distance loss of data without image category labels, L represents the image category label set, d(·) is the distance measurement formula, and the present invention uses the Euclidean distance formula, d(f iw ,c l ) represents a sample u without image category label i After weak enhancement, the feature vector f is obtained by the feature extraction network iw and the prototype c of the category to which this sample belongs l The distance, d(f iw ,c l′ ) represents a sample u without image category label i After weak enhancement, the feature vector f is obtained by the feature extraction network iw and the prototype c of the l′th class l′ distance.
9. A semi-supervised small sample SAR image recognition method, Features: It includes the following steps: The SAR image training data is enhanced in batches, and the enhanced SAR image training data is transmitted to the SAR image recognition model composed of a feature extraction network and a classifier, and the loss of the SAR image recognition model is calculated according to the loss function to complete the training of the SAR image recognition model; The SAR image to be identified is transmitted to the trained SAR image recognition model for SAR image recognition.
10. A computer program product comprising a computer program, It is characterized in that The computer program implements the steps of the method as claimed in claim 9 when executed by a processor.
Citation Information
Patent Citations
SAR image target recognition method based on semi-supervision
CN115775353A
Remote sensing image cross-domain small sample classification method based on confidence coefficient dynamic learning
CN117152606A
Cited By
Semi-supervised image recognition method based on adaptive contrast learning
CN119478591A
Unsupervised SAR change detection method based on collaborative change consistency network model
CN119942213A
Unsupervised sar change detection method based on cooperative change consistency network model
CN119942213B
SAR image road segmentation method based on semi-supervised prototype representation learning and pseudo label balance
CN120032130A
General SAR target identification method and system based on end-to-end structure
CN120107811A
