An incremental learning method and apparatus

Abstract

This invention relates to an incremental learning method and apparatus, comprising: acquiring CNN features of a SAR amplitude image and an estimated parameter set of echo data corresponding to the SAR amplitude image; converting the estimated parameter set into electromagnetic scattering center features; obtaining fused features based on the CNN features and the electromagnetic scattering center features; performing regularization processing; constructing a loss function based on the regularized fused features; and incrementally learning a pre-constructed SCF-NET using the loss function and normalized weights until the pre-constructed SCF-NET reaches a preset convergence condition, thereby obtaining the incrementally learned SCF-NET. This improves the accuracy and stability of incremental learning, and the resulting incrementally learned SCF-NET exhibits high flexibility, scalability, and practicality.

Description

Technical Field

[0001] This invention relates to the field of incremental learning technology, and more specifically, to an incremental learning method and apparatus. Background Technology

[0002] With the rapid development of deep learning technology, the application of deep neural networks in remote sensing target classification has become increasingly widespread and mature. However, most of these network models are trained offline, making it relatively difficult to adjust their classification capabilities to adapt to new categories once training is complete. This challenge is particularly pronounced for SAR (Synthetic Aperture Radar) target recognition systems, as they require continuous collection and integration of new data to ensure the network can be updated in a timely manner and reflect environmental changes.

[0003] After the existing training process is completed, the network model faces the challenge of being unable to adapt to future data. When the system receives new, previously unseen category data, in order to maintain or improve classification performance, it often needs to retrain the network comprehensively using all accumulated data. In addition, in order to preserve historical data, the system also needs to be equipped with a large amount of storage space, which undoubtedly increases the burden on hardware resources.

[0004] Due to the aforementioned limitations, traditional classification methods exhibit significant limitations when facing application scenarios requiring an ever-expanding number of categories. Therefore, in order to meet the growing demand for category recognition and overcome the shortcomings of existing technologies, there is an urgent need to develop new class incremental learning methods to improve the flexibility, scalability, and practicality of SAR target recognition systems. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, the present invention provides an incremental learning method and apparatus.

[0006] According to a first aspect of the present invention, an incremental learning method is provided, the method comprising:

[0007] Obtain the CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image;

[0008] The estimated parameter set is converted into electromagnetic scattering center features;

[0009] The CNN features and the electromagnetic scattering center features are fused using a pre-constructed SCF-NET to obtain fused features;

[0010] The fused features are regularized, and a loss function is constructed based on the regularized fused features.

[0011] When performing incremental learning on the pre-built SCF-NET, the loss function is used to optimize the pre-built SCF-NET after adding a new category, and the normalized weight of the new category is obtained according to the new category and the original category of the SCF-NET. The pre-built SCF-NET is then subjected to the next incremental learning based on the normalized weight of the new category until the pre-built SCF-NET reaches the preset convergence condition, thus obtaining the incrementally learned SCF-NET.

[0012] Optionally, the step of obtaining the CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image includes:

[0013] CNN features of the SAR amplitude image are extracted using a simplified VGG network;

[0014] The estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained by using a simplified ASC scattering model.

[0015] Optionally, the estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained according to the following formula:

[0016]

[0017] in, This represents the set of estimated parameters corresponding to the i-th electromagnetic scattering center in the estimated SAR amplitude image. Φ represents the value when the formula in parentheses reaches its minimum. i , Φ i This represents the parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image. A represents the vector information obtained by converting the raw echo data corresponding to the SAR amplitude image. i This represents the amplitude value of the i-th electromagnetic scattering center. This represents the normalized echo data of the i-th electromagnetic scattering center.

[0018] Optionally, the fusion feature is represented as follows:

[0019]

[0020] in, This indicates the fusion feature. This refers to the fully connected layer in the SCF-NET. This represents the similarity matrix after normalization. This indicates the characteristics of the electromagnetic scattering center.

[0021] Optionally, the normalized similarity matrix is ​​represented as follows:

[0022]

[0023] Where softmax(·) represents the softmax function, M s Let d represent a similar matrix. k The CNN features representing the SAR amplitude image and the electromagnetic scattering center features Dimensions.

[0024] Optionally, the loss function is expressed as follows:

[0025] l=(1-λ)l ac +λl ce ;

[0026] Where l represents the loss function, l ac Let l represent the angle constraint loss function. ce Let λ represent the cross-entropy loss function, and let λ represent the preset weights.

[0027] Optionally, the angle constraint loss function is expressed as follows:

[0028]

[0029] in, This represents the weight vector in the k-th row of the weight matrix of the classifier in SCF-NET. Let ||||2 represent the fusion feature, and let L2 norm represent the L2 norm.

[0030] Optionally, the cross-entropy loss function is expressed as follows:

[0031]

[0032] Where N represents the number of targets in the SAR amplitude image, M represents the total number of target categories, and t i'c y represents the probability that target i' is correctly classified. i'c This represents the symbolic function corresponding to target i'.

[0033] According to a second aspect of the present invention, an incremental learning apparatus is provided, the apparatus comprising:

[0034] The extraction module is used to obtain the CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image;

[0035] A conversion module is used to convert the estimated parameter set into electromagnetic scattering center features;

[0036] The fusion module is used to fuse the CNN features and the electromagnetic scattering center features using a pre-built SCF-NET to obtain fused features;

[0037] A construction module is used to perform regularization processing on the fused features and construct a loss function based on the regularized fused features;

[0038] The incremental learning module is used to optimize the pre-built SCF-NET after adding a new category by using the loss function when incrementally learning the pre-built SCF-NET, and to obtain the normalized weight of the new category based on the new category and the original category of the SCF-NET, and to perform the next incremental learning on the pre-built SCF-NET based on the normalized weight of the new category, until the pre-built SCF-NET reaches the preset convergence condition, thus obtaining the incrementally learned SCF-NET.

[0039] The technical solution provided by this invention may include the following beneficial effects:

[0040] The above technical solution addresses the growing demand for category recognition by incrementally learning SCF-NET. By utilizing the fusion features related to the electromagnetic scattering center of the target in the SAR amplitude image to construct the loss function, the accuracy and stability of incremental learning are effectively improved. The resulting incrementally learned SCF-NET has high flexibility, scalability, and practicality.

[0041] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a flowchart illustrating an incremental learning method according to an exemplary embodiment.

[0044] Figure 2 This is a schematic diagram illustrating the characteristic transformation of an electromagnetic scattering center according to an exemplary embodiment.

[0045] Figure 3 This is a schematic diagram illustrating a fusion feature acquisition process according to an exemplary embodiment.

[0046] Figure 4 This is a pre-built network architecture diagram of SCF-NET illustrated according to an exemplary embodiment.

[0047] Figure 5 This is a pre-built network architecture diagram of SCF-NET illustrated according to an exemplary embodiment.

[0048] Figure 6 This is a diagram illustrating the experimental results of electromagnetic scattering center feature ablation according to an exemplary embodiment.

[0049] Figure 7 This is an ablation experiment result diagram illustrating a cross-attention feature fusion method according to an exemplary embodiment.

[0050] Figure 8 This is a graph illustrating the ablation experiment analysis results of an overfitting strategy according to an exemplary embodiment.

[0051] Figure 9 This is a graph illustrating the ablation experiment analysis results of another overfitting strategy according to an exemplary embodiment.

[0052] Figure 10 This is a graph illustrating the ablation experiment analysis results of another overfitting strategy according to an exemplary embodiment.

[0053] Figure 11 This is a diagram illustrating the ablation experiment analysis results of an angle-constrained loss function according to an exemplary embodiment. Detailed Implementation

[0054] Figure 1 This is a flowchart illustrating an incremental learning method according to an exemplary embodiment, such as... Figure 1 As shown, the method includes the following steps.

[0055] S101. Obtain the CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image.

[0056] Optionally, S101 may include:

[0057] CNN features of SAR amplitude images are obtained using a simplified VGG network;

[0058] The estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained by using a simplified ASC scattering model.

[0059] It is understandable that when extracting CNN (Convolutional Neural Network) features from SAR amplitude images, the simplified VGG16 network feature extractor constructed in this invention is used. The original VGG16 network has two conv3-512 layers, one conv3-256 layer, and all fully connected layers removed, so as to obtain CNN features from SAR amplitude images with less computation.

[0060] Optionally, the estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained according to the following formula:

[0061]

[0062] in, This represents the set of estimated parameters corresponding to the i-th electromagnetic scattering center in the estimated SAR amplitude image. Φ represents the value when the formula in parentheses reaches its minimum. i , Φ i This represents the parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image. A represents the vector information obtained by converting the raw echo data corresponding to the SAR amplitude image. i This represents the amplitude value of the i-th electromagnetic scattering center. This represents the normalized echo data of the i-th electromagnetic scattering center.

[0063] It is worth mentioning that the normalized echo data of the i-th electromagnetic scattering center It can be represented as follows:

[0064]

[0065] Among them, L i It is the length of the electromagnetic scattering center i, c,f and These represent the speed of electromagnetic wave propagation, the radar frequency, and the azimuth observation angle, respectively. The initial pointing angle of the electromagnetic scattering center i is represented by x. i y i Let be the azimuth and range positions of the electromagnetic scattering center i, and exp be an exponential function. After iterative optimization, the estimated parameter set of the electromagnetic scattering center i can be obtained. Where “^” represents the approximate result of the optimization, and Let represent the coordinates of the i-th electromagnetic scattering center. Assuming each target class contains n samples, the size of the estimated parameter set is n×N. sc ×6, N sc This represents the number of scattering centers to be screened in a single sample; and the parameters are set together. Divided into and and These are the real and imaginary parts of the amplitude at the scattering center, respectively.

[0066] S102. Convert the estimated parameter set into electromagnetic scattering center features.

[0067] Understandable Figure 2 This is a schematic diagram illustrating the characteristic transformation of an electromagnetic scattering center according to an exemplary embodiment, with reference to... Figure 2 The visual bag-of-words model is used to transform the scattering center parameter set of each target into an electromagnetic scattering center feature vector. The k-means clustering algorithm is then used to search for the estimated parameter set. Cluster centers, multiple cluster centers form a visual vocabulary. For measured SAR complex data, assuming the number of extracted electromagnetic scattering centers is M', these M' electromagnetic scattering centers are assigned to N' different cluster centers. The number of electromagnetic scattering centers assigned to each cluster center is counted and regularized, transforming the input SAR data into an N'-dimensional vector. For this feature vector, targets of the same category should follow the same or similar frequency distribution. Thus, each target in the SAR amplitude image is represented as an electromagnetic scattering center feature vector. For example, the clustering is based on Φ... i Clustering is performed, and each target may have multiple electromagnetic scattering centers. The Φ values ​​of all electromagnetic scattering centers of all targets are then used to classify the data. i Clustering is performed, resulting in cluster centers. In other words, each cluster center is actually... In this way, they are clustered into N' categories, which are N' cluster centers. The N' cluster centers are used to form a visual vocabulary. Assign each scattering center of the target to a cluster center according to the nearest neighbor rule, if... Allocate 1, If we assign 2 to each target and 0 to the others, then the feature vector of this target is {1,2,…,0}.

[0068] S103. The CNN features and electromagnetic scattering center features are fused using the pre-constructed SCF-NET to obtain fused features.

[0069] It is worth mentioning, and understandably so, Figure 3 This is a schematic diagram illustrating a fusion feature acquisition process according to an exemplary embodiment. See also: Figure 3 Based on the cross-attention model, CNN features are... Set as the query vector Q to include the electromagnetic scattering center features Let K be the key vector. Since both features mentioned above are extracted from the same echo data, the similarity matrix can be obtained by vector multiplication.

[0070] The similarity matrix is ​​represented as follows:

[0071]

[0072] in,() TIt is a vector transpose operation.

[0073] Alternatively, the normalized similarity matrix can be represented as follows:

[0074]

[0075] Where softmax(·) represents the softmax function, M s Let d represent a similar matrix. k CNN features and electromagnetic scattering center features representing SAR amplitude images The dimensions of the query vector Q and the key vector K are the same in this invention, and both are 256.

[0076] Optionally, the electromagnetic scattering center characteristics are... The value vector V is multiplied by the similarity matrix, and then passed through a fully connected layer to calculate the fused features. The similarity matrix M... s Not only used to describe CNN features Characteristics of electromagnetic scattering centers The degree of similarity can be used as a weight for fusing feature vectors. To make the electromagnetic scattering center features... It plays a major role in the final fusion feature, and is used as the value vector V and similarity matrix M. s Multiplication can be used to represent the fused features as follows:

[0077]

[0078] in, Indicates fusion features, This represents the fully connected layer in SCF-NET. This represents the similarity matrix after normalization. The electromagnetic scattering center feature is represented, and its function is to transform the fused features into a 128*1 matrix to improve the network's fitting ability. Therefore, CAF-NET (Feature fusing based on cross attention mechanism) in SCF-NET can fuse the CNN features and electromagnetic scattering features of the image into features input to the classifier.

[0079] Optionally, Figure 4 This is a network architecture diagram of a pre-built SCF-NET illustrated according to an exemplary embodiment, with reference to... Figure 4Removing the dropout layer and batch normalization operation from the classifier in the SCF-NET architecture improves the network's fit to existing classes, increasing classification accuracy in the next stage of incremental learning (CIL). A "late stopping" strategy is employed, where the SCF-NET network is trained for an additional period after achieving a good fit. This enhances the clustering of existing class features in the feature space, reducing the probability of new class features falling into cluster centers of existing features. This improves classification accuracy in subsequent stages of incremental learning without significantly reducing the accuracy of the current stage.

[0080] S104. Regularize the fused features and construct a loss function based on the regularized fused features.

[0081] Understandable Figure 5 This is a network architecture diagram of a pre-built SCF-NET illustrated according to an exemplary embodiment, with reference to... Figure 5 The parameters during the fully connected layer operation are analyzed. The classifier is a fully connected layer, and its operation is as follows:

[0082]

[0083] in, This represents the classification result of the target; W and These represent the weights and biases of the fully connected layer, respectively. Ignoring the classifier's bias, the above equation is transformed into a vector form as follows, which makes the classifier's operation process easier to understand:

[0084]

[0085] in, Let W be the weight vector in the k-th row (1≤k≤H), where 1≤k≤H represents different classes in class H. Let y be the weight of classifier k. k Let k be the output value of classifier k. Then the above formula can be transformed into:

[0086]

[0087] Where θ represents a vector and The angle between them, through the above formula, allows us to analyze the three parameters θ in the incremental learning process. and In incremental learning without accessing visible classes, sample imbalance is unavoidable. A deep neural network is a black box that automatically adjusts its parameters to achieve optimization; therefore, it can be understood through the parameters in the last fully connected layer, i.e., the network will adjust the three parameters θ on the right-hand side of the above equation simultaneously. and Ultimately, this achieves effective identification of new categories while avoiding identification bias caused by catastrophic forgetting. Feature regularization operations can be added during the network's forward propagation process, specifically by fusing features from the cross-attention mechanism. Perform regularization to convert the length of each feature input to the classifier to a unit length:

[0088]

[0089] Optionally, an angle-constrained loss function is constructed based on the fusion features to constrain the angle variation between the weights of the new and old classifiers, as follows:

[0090]

[0091] in, This represents the weight vector in the k-th row of the weight matrix of the classifier in SCF-NET. represents the fusion feature, and ||||2 represents the L2 norm.

[0092] The original cross-entropy loss function can be expressed as:

[0093]

[0094] Where N represents the number of targets in the SAR amplitude image, M represents the total number of target categories, and t i'c y represents the probability that target i' is correctly classified. i'c This represents the symbolic function corresponding to target i'.

[0095] Sign function y i'c The definition is as follows:

[0096]

[0097] Where, label i' The label represents the predicted sample i'.

[0098] Based on the angle constraint loss function and the cross-entropy loss function, the loss function is expressed as follows:

[0099] l=(1-λ)l ac +λl ce ;

[0100] Where l represents the loss function, l ac Let l represent the angle constraint loss function. ce Let λ represent the cross-entropy loss function, and let λ represent the preset weights.

[0101] S105. When performing incremental learning on the pre-built SCF-NET, the loss function is used to optimize the pre-built SCF-NET after adding new categories. The normalized weights of the new categories are obtained based on the new categories and the original categories of the SCF-NET. The pre-built SCF-NET is then subjected to the next incremental learning based on the normalized weights of the new categories until the pre-built SCF-NET reaches the preset convergence condition, thus obtaining the incrementally learned SCF-NET.

[0102] It is worth mentioning that the old class weights are used as a guide to standardize the new class weights. In the above regularization steps, features... The angle θ and amplitude are constrained, thus transferring the impact of sample imbalance on the network classifier to the weights. The size of the class determines the weight of the new class.

[0103]

[0104] Where, N old With N new These represent the number of items in the old class and the number of items in the new class, respectively. and Let represent the average of the old class weight and the new class weight, respectively. Then the normalized weight of the new class is:

[0105]

[0106] This multi-step regularization method can transfer and reduce the impact of sample imbalance during incremental learning, thereby improving the performance of incremental learning.

[0107] Understandably, the preset convergence condition could be that the number of incremental learning rounds reaches a certain amount, or that the value of the loss function is less than a preset threshold, depending on the actual situation.

[0108] In one implementation, the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset is used to verify the effectiveness of the invention. Since the classification performance of SAR target images is easily affected by the elevation angle, a dataset with a 17° elevation angle condition is used for training, and a dataset with a 15° elevation angle condition is used for the inference testing phase. Specific information about this dataset is shown in Table 1.

[0109] Table 1

[0110]

[0111] In the incremental learning experiment, the 10 target classes need to be divided into two parts: one part for basic training and the other part for incremental learning. The basic training class is used to train the network to ensure that it has sufficient feature extraction and classification capabilities. Then, the incremental learning stage is used to evaluate the incremental learning capability of this method. Taking both parts into consideration, 6 classes are selected as the basic training class, and the remaining 4 classes are selected as the incremental learning class. In each incremental learning stage, a new class is used to incrementally train the network, which is divided into 5 stages in total, as shown in Table 1.

[0112] In one implementation, Figure 6 This is an experimental result diagram of electromagnetic scattering center feature ablation according to an exemplary embodiment, for reference. Figure 6 Ablation experiments were conducted to compare the impact of fused scattering center features on algorithm performance. First, VGG16 and SCF-NET networks were used on the MSTAR dataset. Since the feature extractor is based on the VGG16 network, only the experimental results for VGG16 were used to measure the incremental learning performance when only image features were applied. Furthermore, since VGG16 is an end-to-end trained network, it can be optimized using distillation loss.

[0113] Compared to using only CNN features, scattering center fusion features not only significantly improve the accuracy of the incremental learning process but also achieve better classification results in the basic learning stage. In other words, using fused features not only effectively improves the performance of the incremental learning process but also achieves good performance in traditional classification tasks. During the incremental learning process, the most significant difference between the two methods appears in stage 3, where the fusion method achieves a classification accuracy of 84.749%, while the method using VGG16 drops to 47.282%. This is due to the significant differences in images of the MSTAR dataset at different azimuth angles, thus the classification accuracy using only CNN features decreases significantly. The incremental learning process inherently requires a certain degree of completeness in the features extracted by the network. This requirement becomes particularly stringent when processing radar target data, which is more sensitive to azimuth angles than optimal target data. In contrast, scattering center features are more complete and can simultaneously overcome the influence of concave angle variations and the requirements of the incremental learning task. By fusing image features and scattering center features, SCF-NET maintains the learnability of the feature extractor and ensures the stability of features at different incremental learning stages. This experiment verifies the effectiveness of fusing SC features.

[0114] Figure 7 This is an ablation experiment result diagram illustrating a cross-attention feature fusion method according to an exemplary embodiment, such as... Figure 7As shown, the impact of fully connected layer fusion and cross-attention feature fusion on algorithm performance is compared. It can be seen that in the basic training phase, the classification accuracy of the fully connected layer fusion method is 95.450%, while the accuracy of the cross-attention feature fusion method is 98.082%. This indicates that relying solely on fully connected layers for feature fusion is not the optimal method for integrating these two features. In terms of performance across all stages of incremental learning, the cross-attention fusion method of SCF-NET consistently outperforms the fully connected method. This demonstrates that the cross-attention method effectively utilizes the stability of the scattering center features.

[0115] Figure 8 This is a graph illustrating the ablation experiment analysis results of an overfitting strategy according to an exemplary embodiment, such as... Figure 8 As shown, this part of the experiment analyzes the impact of overfitting on feature clustering by measuring the distance variance of the output features. For features belonging to the same class... Two characteristics and The distance between them is:

[0116]

[0117] Among them ||·|| Euclidean The representative value is Euclidean distance. Since incremental learning focuses more on the volatility of feature clustering, the average standard deviation of the features for each training epoch is expressed as follows:

[0118]

[0119] Where, N CLASS Represents the number of categories, while Representatives are examples The number can be obtained from the above formula. Figure 8 This shows the difference between having and not having dropout layers at different training stages.

[0120] based on Figure 8 The two accuracy trends show that the network without dropout layers fits faster than the network with dropout layers. Then, looking at the standard deviation trends, the standard deviation of the network without dropout layers is consistently lower than that of the network with dropout layers. This result verifies that removing the dropout layer structure helps improve the stability of feature clustering. Furthermore, as accuracy rapidly increases, the standard deviation also shows a rapid increase. Subsequently, the standard deviation gradually decreases as training progresses, indicating that the late stopping strategy also helps improve the stability of feature clustering.

[0121] Figure 9This is an ablation experiment analysis result diagram illustrating another overfitting strategy according to an exemplary embodiment, showing network training results with and without dropout layers. Figure 9 As shown. Figure 9 The conclusion and Figure 8 Similarly, removing the dropout layer and the late stopping strategy have a more significant impact on feature clustering. These experimental results validate the effectiveness of our two methods from the perspectives of cluster stability and density.

[0122] To discuss the impact of overfitting methods on incremental learning performance, this section uses four methods as shown in Table 2:

[0123] Table 2

[0124]

[0125] Figure 10 This is a graph illustrating the ablation experiment analysis results of another overfitting strategy according to an exemplary embodiment. Figure 10 It can be seen that removing the dropout layer and the late stopping strategy both improved the performance in the incremental learning stage to varying degrees, with the late stopping strategy showing the smallest improvement. It is noteworthy that the highest classification accuracy in the basic stage reached 99.147% when using the initial network model. However, the best-performing blue line in the incremental learning stage achieved 98.540%. In other words, the overfitting method sacrifices some recognition performance in the basic stage in exchange for improved performance in the incremental learning stage.

[0126] As can be seen from the four methods above, simultaneously removing the dropout layer and employing a late-stopping strategy yields only the smallest improvement, with accuracies of 88.955% and 84.920% in the first and second stages, respectively, compared to 88.578% and 84.662% for the initial method. However, the improvement from simply removing the dropout layer is significant, achieving accuracies of 93.099% and 87.827% in the first and second stages, respectively. This is because the dropout layer slows down the network fitting process by potentially inactivating neurons, which is more effective than the late-stopping strategy. However, comparing the difference between the blue and yellow lines with the difference between the green and red lines reveals that the performance improvement from the late-stopping strategy varies under different conditions. This is due to the randomness introduced by the dropout layer into the training process, which significantly impacts the network's learning process.

[0127] For the angle-constrained loss function proposed in this paper, the following experiments were conducted to verify its effectiveness in restricting angles and improving incremental learning performance during network training. This part mainly compares two methods: one using only the cross-entropy loss function and the other using both cross-entropy and angle-constrained loss functions. For the second case, λ was set to 0.5, and the same hyperparameters, such as the learning rate and optimizer, were used in the training process for both methods. Considering that the value of the loss is difficult to control during the learning process, we also compared the cosine value of the angle between the new class features and the old classifier weights when the network training reached different accuracies. This is referred to as the sampling accuracy in this paper, including 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 0.95.

[0128] Table 3

[0129] Sampling accuracy 0.4 0.5 0.6 0.7 0.8 0.9 0.95 CEloss 0.21 0.18 0.16 0.14 0.16 0.13 0.11 CAloss+CEloss 0.16 0.14 0.14 0.15 0.15 0.12 0.09

[0130] The experimental results are shown in Table 3. It is important to note that since the network training progress is difficult to manually intervene in, the accuracy is first verified in the testing phase after each training stage to determine if it exceeds the sampling accuracy, thus obtaining the corresponding experimental results. However, for a sampling accuracy of 0.7, the accuracies of the two networks differ significantly, at 0.704 and 0.749 respectively. This inevitably leads to a large error in the experimental results at a sampling accuracy of 0.7. Table 3 concludes that when the network trained with the angle-constrained loss function reaches a certain accuracy, the angle between the new feature vector and the old classifier weights is smaller than that of the network without the angle-constrained loss function, indicating that it is closer to orthogonality. This proves that during network optimization, our angle-constrained loss function has a certain constraint effect on the angle.

[0131] Figure 11 This is a graph illustrating the ablation experiment analysis results of an angle-constrained loss function according to an exemplary embodiment. See also... Figure 11 This section also compares the impact of the angle constraint loss function on incremental learning. At each stage of the incremental learning process, the network trained with the angle constraint loss function performs slightly better than the network without it. The largest difference is observed in stage 4, with accuracies of 57.019% and 65.151% respectively, a difference of 8.13%. This part of the experiment demonstrates that introducing the angle constraint loss function can improve the performance of incremental learning.

[0132] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0133] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0134] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. An incremental learning method, characterized in that, The method includes: Extract the CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image; The estimated parameter set is converted into electromagnetic scattering center features; The CNN features and the electromagnetic scattering center features are fused using a pre-constructed SCF-NET to obtain fused features; The fused features are regularized, and a loss function is constructed based on the regularized fused features. When performing incremental learning on the pre-built SCF-NET, the loss function is used to optimize the pre-built SCF-NET after adding a new category, and the normalized weight of the new category is obtained according to the new category and the original category of the SCF-NET. The pre-built SCF-NET is then subjected to the next incremental learning based on the normalized weight of the new category until the pre-built SCF-NET reaches the preset convergence condition, thus obtaining the incrementally learned SCF-NET.

2. The incremental learning method according to claim 1, characterized in that, The extracted CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image include: CNN features of the SAR amplitude image are extracted using a simplified VGG network; The estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained by using a simplified ASC scattering model.

3. The incremental learning method according to claim 2, characterized in that, The estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained according to the following formula: in, This represents the set of estimated parameters corresponding to the i-th electromagnetic scattering center in the estimated SAR amplitude image. Φ represents the value when the formula in parentheses reaches its minimum. i Φ i This represents the parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image. A represents the vector information obtained by converting the raw echo data corresponding to the SAR amplitude image. i This represents the amplitude value of the i-th electromagnetic scattering center. This represents the normalized echo data of the i-th electromagnetic scattering center.

4. The incremental learning method according to claim 1, characterized in that, The fusion features are represented as follows: in, This indicates the fusion feature. This refers to the fully connected layer in the SCF-NET. This represents the similarity matrix after normalization. This indicates the characteristics of the electromagnetic scattering center.

5. The incremental learning method according to claim 4, characterized in that, The normalized similarity matrix is ​​represented as follows: Where softmax(·) represents the softmax function, M s Let d represent a similar matrix. k The CNN features representing the SAR amplitude image and the electromagnetic scattering center features Dimensions.

6. The incremental learning method according to claim 1, characterized in that, The loss function is expressed as follows: l=(1-λ)l ac +λl ce ; Where l represents the loss function, l ac Let l represent the angle constraint loss function. ce Let λ represent the cross-entropy loss function, and let λ represent the preset weights.

7. The incremental learning method according to claim 6, characterized in that, The angle constraint loss function is expressed as follows: in, This represents the weight vector in the k-th row of the weight matrix of the classifier in SCF-NET. Let ||||2 represent the fusion feature, and let L2 norm represent the L2 norm.

8. The incremental learning method according to claim 6, characterized in that, The cross-entropy loss function is expressed as follows: Where N represents the number of targets in the SAR amplitude image, M represents the total number of target categories, and t i'c y represents the probability that target i' is correctly classified. i'c This represents the symbolic function corresponding to target i'.

9. An incremental learning device, characterized in that, The device includes: The extraction module is used to obtain the CNN features of the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image; A conversion module is used to convert the estimated parameter set into electromagnetic scattering center features; The fusion module is used to fuse the CNN features and the electromagnetic scattering center features using a pre-built SCF-NET to obtain fused features; A construction module is used to perform regularization processing on the fused features and construct a loss function based on the regularized fused features; The incremental learning module is used to optimize the pre-built SCF-NET after adding a new category by using the loss function when incrementally learning the pre-built SCF-NET, and to obtain the normalized weight of the new category based on the new category and the original category of the SCF-NET, and to perform the next incremental learning on the pre-built SCF-NET based on the normalized weight of the new category, until the pre-built SCF-NET reaches the preset convergence condition, thus obtaining the incrementally learned SCF-NET.

Images