Incremental learning method and device

Abstract

The invention relates to an incremental learning method and device. The method comprises the steps of obtaining CNN features of an SAR amplitude image and an estimation parameter set of echo data corresponding to the SAR amplitude image; converting the estimation parameter set into electromagnetic scattering center characteristics; obtaining fusion features according to the CNN features and the electromagnetic scattering center features; carrying out regularization processing, and constructing a loss function according to the fused features after regularization processing; and performing incremental learning on the pre-constructed SCF-NET by using the loss function and the normalized weight until the pre-constructed SCF-NET reaches a preset convergence condition, and obtaining the SCF-NET after incremental learning. The accuracy and stability of incremental learning are improved, and the SCF-NET after incremental learning has high flexibility, expandability and practicability.

Description

Technical Field

[0001] The present invention relates to the technical field of incremental learning, and in particular to an incremental learning method and device. Background Art

[0002] With the rapid development of deep learning technology, the application of deep neural networks in the field of remote sensing target classification has become increasingly widespread and mature. Nevertheless, most of these network models are trained in an offline environment. Once the training is completed, it is relatively difficult to adjust their classification capabilities to adapt to new categories. This challenge is particularly prominent for SAR (Synthetic Aperture Radar) target recognition systems because they need to continuously collect and integrate new data to ensure that the network can be updated in a timely manner and reflect changes in the environment.

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

[0004] Due to the above limitations, traditional classification methods have shown obvious limitations when facing application scenarios that require the continuous expansion of the number of categories. Therefore, in order to meet the growing demand for category recognition and overcome the shortcomings of existing technologies, it is urgent to develop new incremental learning methods to improve the flexibility, scalability and practicality of SAR target recognition systems. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an incremental learning method and device.

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

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

[0008] converting the estimated parameter set into electromagnetic scattering center features;

[0009] Using the pre-built SCF-NET to fuse the CNN features and the electromagnetic scattering center features to obtain fused features;

[0010] Regularizing the fused features, and constructing a loss function according to the regularized fused features;

[0011] When performing incremental learning on the pre-constructed SCF-NET, the pre-constructed SCF-NET after adding the new category is optimized using the loss function, and the normalized weight of the new category is obtained according to the new category and the original category of the SCF-NET, and the pre-constructed SCF-NET is subjected to the next incremental learning according to the normalized weight of the new category, until the pre-constructed SCF-NET reaches a preset convergence condition, thereby obtaining the SCF-NET after incremental learning.

[0012] Optionally, the acquiring of CNN features of the SAR amplitude image and an estimation parameter set of echo data corresponding to the SAR amplitude image comprises:

[0013] Extracting CNN features of the SAR amplitude image through a simplified VGG network;

[0014] The estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained through the 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, represents the estimated parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image, Indicates Φ when the formula in brackets reaches its minimum value i , Φ i represents the parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image, Represents the vector information obtained by converting the echo raw data corresponding to the SAR amplitude image, A i represents the amplitude value of the i-th electromagnetic scattering center, Represents the normalized echo data of the i-th electromagnetic scattering center.

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

[0019]

[0020] in, represents the fusion feature, represents the fully connected layer in the SCF-NET, represents the normalized similarity matrix, Represents the electromagnetic scattering center feature.

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

[0022]

[0023] Among them, softmax(·) represents the softmax function, M s represents the similarity matrix, d k Representing the CNN features of the SAR amplitude image and the electromagnetic scattering center features Dimension.

[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 represents the angle constraint loss function, l ce represents the cross entropy loss function, and λ represents the preset weight.

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

[0028]

[0029] in, represents the kth row weight vector in the weight matrix of the classifier in SCF-NET, represents the fusion feature, and ||||2 represents 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 represents the probability that target i' is correctly classified, y i'c Represents the symbolic function corresponding to target i'.

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

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

[0035] A conversion module, used for converting the estimated parameter set into electromagnetic scattering center features;

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

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

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

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

[0040] Through the above technical scheme, SCF-NET is incrementally learned to cope with the growing demand for category recognition. The loss function is constructed by using the fusion features related to the electromagnetic scattering center of the target in the SAR amplitude image, which effectively improves the accuracy and stability of incremental learning. The SCF-NET obtained after incremental learning 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention but do not constitute a limitation of the present invention. In the accompanying drawings:

[0043] Figure 1 The figure is a flow chart of an incremental learning method according to an exemplary embodiment.

[0044] Figure 2 It is a schematic diagram showing a characteristic conversion of an electromagnetic scattering center according to an exemplary embodiment.

[0045] Figure 3 The figure is a schematic diagram of a fusion feature acquisition process according to an exemplary embodiment.

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

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

[0048] Figure 6 The figure is a result diagram of an electromagnetic scattering center feature ablation experiment according to an exemplary embodiment.

[0049] Figure 7 It is a diagram of ablation experiment results of a cross-attention feature fusion method according to an exemplary embodiment.

[0050] Figure 8 is a diagram showing analysis results of an ablation experiment of an overfitting strategy according to an exemplary embodiment.

[0051] Fig. 9 is a diagram showing analysis results of an ablation experiment of another overfitting strategy according to an exemplary embodiment.

[0052] Fig.10 is a diagram of analysis results of an ablation experiment of yet another overfitting strategy according to an exemplary embodiment.

[0053] Fig.11 It is a diagram of analysis results of an ablation experiment of an angle constraint loss function according to an exemplary embodiment. DETAILED DESCRIPTION

[0054] Figure 1 is a flow chart of an incremental learning method according to an exemplary embodiment. Figure 1 As shown, the method includes the following steps.

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

[0056] Optionally, S101 may include:

[0057] The CNN features of the SAR amplitude image are obtained through the simplified VGG network;

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

[0059] It can be understood that when extracting the CNN (Convolutional Neural Network) features of the SAR amplitude image, the simplified VGG16 network feature extractor constructed by the present invention is used, and two conv3-512 layers, one conv3-256 layer and all fully connected layers of the original VGG16 network are removed to obtain the CNN features of the SAR amplitude image with less calculation amount.

[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, represents the estimated parameter set corresponding to the i-th electromagnetic scattering center in the estimated SAR amplitude image, Indicates Φ when the formula in brackets reaches its minimum value i , Φ i represents the parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image, Represents the vector information obtained by converting the echo raw data corresponding to the SAR amplitude image, A i represents the amplitude value of the i-th electromagnetic scattering center, 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 expressed as follows:

[0064]

[0065] Among them, L i is the length of the electromagnetic scattering center i, c,f and They represent the electromagnetic wave propagation speed, radar frequency and azimuth observation angle respectively. represents the initial pointing angle of the electromagnetic scattering center i, x i ,y i is the azimuth and distance position of the electromagnetic scattering center i, and exp is an exponential function. After iterative optimization, the estimated parameter set of the electromagnetic scattering center i can be obtained: Where “^” represents the approximate results of optimization, and Represents the coordinate position of the i-th electromagnetic scattering center. Assuming that each type of target contains n samples, the size of the estimated parameter set is n×N sc ×6,N sc Indicates the number of scattering centers that need to be screened in a single sample; and the parameters are concentrated Divide into and and are the real and imaginary parts of the amplitude at the scattering center, respectively.

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

[0067] Understandably, Figure 2 is a schematic diagram showing a conversion of electromagnetic scattering center characteristics according to an exemplary embodiment, with reference to Figure 2 , the visual word bag model is used to convert the scattering center parameter set of each target into an electromagnetic scattering center feature vector. The k-means clustering algorithm is used to search for the estimated parameter set The cluster centers of multiple cluster centers form a visual vocabulary For the measured SAR complex data, assume that the number of extracted electromagnetic scattering centers is M', and 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 regularization is performed to convert the input SAR data into an N'-dimensional vector. For this feature vector, targets of the same category should obey 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, clustering is a pair of Φ i Clustering is performed. Each target may have multiple electromagnetic scattering centers. The Φ of all electromagnetic scattering centers of all targets is clustered. i Clustering is performed to form cluster centers. In other words, each cluster center is actually In the form of N' categories, that is, N' cluster centers N' cluster centers are combined into a visual vocabulary According to the rule of the closest distance, each scattering center of the target is assigned to the cluster center. If Allocate 1, Assign 2 and the others 0, then the feature vector of this target is {1,2,…,0}.

[0068] S103. Use the pre-built SCF-NET to fuse the CNN features and the electromagnetic scattering center features to obtain fused features.

[0069] It is worth mentioning that, understandably, Figure 3 is a schematic diagram of a fusion feature acquisition process according to an exemplary embodiment. Figure 3 , based on the cross attention model, the CNN features Set as query vector Q, and transform the electromagnetic scattering center feature Set as the key vector K, the above two features are extracted from the same echo data, so the similarity matrix can be obtained by vector multiplication.

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

[0071]

[0072] in,() Tis the vector transpose operation.

[0073] Optionally, the normalized similarity matrix is ​​expressed as follows:

[0074]

[0075] Among them, softmax(·) represents the softmax function, M s represents the similarity matrix, d k CNN features representing SAR amplitude images and electromagnetic scattering center features The dimension of , that is, the dimension of the query vector Q and the key vector K, in the present invention, the dimension of the query vector Q and the key vector K is the same and both are 256.

[0076] Optionally, let the electromagnetic scattering center feature As the value vector V, it is multiplied by the similarity matrix and passed through the fully connected layer to calculate the fusion feature. s Not only used to describe CNN features Electromagnetic scattering center characteristics The similarity of the electromagnetic scattering center features can be used as the weight of the fusion feature vector. It plays a major role in the final fusion feature, making it the value vector V and the similarity matrix M s Multiplying together, the fused features can be expressed as follows:

[0077]

[0078] in, represents the fusion feature, represents the fully connected layer in SCF-NET, represents the normalized similarity matrix, It represents the electromagnetic scattering center feature, which is used to convert the fusion feature into a 128*1 matrix to improve the network's fitting ability. Therefore, CAF-NET (Featurefusing 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 is a network architecture diagram of a pre-built SCF-NET according to an exemplary embodiment, referring to Figure 4, removing the dropout layer and batch normalization operation in the classifier in the SCF-NET structure can improve the network's fit to the seen classes and increase the classification accuracy of the network in the next stage of incremental learning (CIL). The "late stopping" strategy is adopted, that is, after the SCF-NET network is well fitted, it is properly trained for a certain period of time to improve the clustering degree of the seen class features in the feature space and reduce the probability that the new class features fall into the cluster center of the seen class features. The classification accuracy of the subsequent stage of incremental learning is improved without significantly reducing the accuracy of the current stage.

[0080] S104, performing regularization processing on the fused features, and constructing a loss function according to the fused features after the regularization processing.

[0081] Understandably, Figure 5 is a network architecture diagram of a pre-built SCF-NET according to an exemplary embodiment, referring to Figure 5 , analyze the parameters in the fully connected layer operation process. The classifier is a fully connected layer, and its operation is as follows:

[0082]

[0083] in, Represents the classification result of the target; W and Represent the weight and bias of the fully connected layer respectively. Ignoring the bias of the classifier, convert the above formula into the vector form as follows, which makes it easier to understand the operation process of the classifier:

[0084]

[0085] in, is the k-th row (1≤k≤H) weight vector of W, where 1≤k≤H represents different classes in H. Represented as the weight of classifier k, y k Expressed as the output value of classifier k, the above formula is converted to:

[0086]

[0087] Among them, θ represents the vector and The angle between them can be analyzed through the above formula respectively. The three parameters θ, and In the incremental learning process without touching the visible class, sample imbalance is inevitable. A deep neural network is a black box that automatically adjusts parameters to achieve optimization. Therefore, the parameters in the last fully connected layer can be adjusted by the network through the simultaneous adjustment of the three parameters θ, and Finally, it is possible to effectively identify new categories while avoiding recognition bias caused by catastrophic forgetting. Feature regularization operations can be added during the network forward propagation process, that is, the fusion features of the cross-attention mechanism Regularization is performed so that the length of each feature input to the classifier is converted to unit length:

[0088]

[0089] Optionally, an angle constraint loss function for constraining the angle change between the weights of the new and old category classifiers is constructed based on the fusion features, which is expressed as follows:

[0090]

[0091] in, represents the kth row weight vector in the weight matrix of the classifier in SCF-NET, represents fusion features, 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 represents the probability that target i' is correctly classified, y i'c Represents the symbolic function corresponding to target i'.

[0095] Symbolic function y i'c is defined as follows:

[0096]

[0097] Among them, label i' Represents the predicted label of sample i'.

[0098] According to 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] Among them, l represents the loss function, l ac represents the angle constraint loss function, l ce represents the cross entropy loss function, and λ represents the preset weight.

[0101] S105. When performing incremental learning on the pre-constructed SCF-NET, the pre-constructed SCF-NET after adding the new category is optimized using the loss function, and the normalized weight of the new category is obtained according to the new category and the original category of the SCF-NET, and the pre-constructed SCF-NET is subjected to the next incremental learning according to the normalized weight of the new category, until the pre-constructed SCF-NET reaches the preset convergence condition, and the SCF-NET after incremental learning is obtained.

[0102] It is worth mentioning that the weight of the old class is used as a guide to normalize the weight of the new class. The angle θ and amplitude of are constrained, thereby transferring the impact of sample imbalance on the network classifier to the weights The size of , thus normalizing the weight of the new class:

[0103]

[0104] Among them, N old With N new Represent the number of old classes and the number of new classes respectively, and 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] Through this multi-step regularization method, the impact of sample imbalance in the incremental learning process can be transferred and reduced, thereby improving the performance of incremental learning.

[0107] It is understandable that the preset convergence condition may be that the number of incremental learning rounds reaches a certain number, or that the value of the loss function is less than a preset threshold, which is determined based on actual conditions.

[0108] In one embodiment, the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset is used to verify the effectiveness of the present invention. Since the classification performance of SAR target images is easily affected by the pitch angle, the dataset with a pitch angle of 17° is used for training, and the dataset with a pitch angle of 15° is used for the inference test phase. The specific information of the dataset is shown in Table 1.

[0109] Table 1

[0110]

[0111] In the incremental learning experiment, these 10 categories of targets need to be divided into two parts: one part is used for basic training and the other part is used 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 ability of this method. Taking these two parts into consideration, 6 categories are selected as basic training classes, and the remaining 4 categories are selected as incremental learning classes. 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 embodiment, Figure 6 is a diagram showing the results of an electromagnetic scattering center feature ablation experiment according to an exemplary embodiment, referring to Figure 6 , an ablation experiment was conducted. In order to compare the impact of fusing scatter center features on algorithm performance, the VGG16 network and SCF-NET were first used on the MSTAR dataset. Since the feature extractor is based on the VGG16 network, only the experimental results of VGG16 are used to measure the incremental learning performance when only image features are applied. In addition, since VGG16 is an end-to-end training network, it can be optimized using distillation loss.

[0113] Compared with using only CNN features, the scattering center fusion features not only greatly improve the accuracy of the incremental learning process, but also obtain better classification results in the basic learning stage. That is, using fused features can not only effectively improve the performance of the incremental learning process, but also achieve better performance in traditional classification tasks. In the incremental learning process, the most significant difference between the two methods occurs in stage 3, where the classification accuracy reaches 84.749%, while the method using VGG16 drops to 47.282%. This is caused by the huge difference in images at different azimuths in the MSTAR dataset, so the classification accuracy using only CNN features decreases significantly. The incremental learning process inherently requires that the features extracted by the network have a certain degree of integrity. This requirement becomes particularly strict when processing radar target data, which is more sensitive to azimuth than the optimal target data. In contrast, the scattering center features are more complete and can simultaneously overcome the influence of the change in the depression angle 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 the features at different incremental learning stages. This experiment verifies the effectiveness of the fused SC features.

[0114] Figure 7 is a diagram showing a result of an ablation experiment of a cross-attention feature fusion method according to an exemplary embodiment. Figure 7As shown in the figure, the impact of the method of fusing multiple feature vectors by fully connected layer fusion and cross-attention feature fusion on the algorithm performance is compared. It can be seen that in the basic training stage, the classification accuracy of the fully connected fusion feature method is 95.450%, while the accuracy of the cross-attention feature fusion method is 98.082%. This shows that relying solely on the fully connected layer for feature fusion is not the best way to integrate the two features. In terms of performance at all stages of incremental learning, the cross-attention fusion method of SCF-NET is always better than the fully connected method. This shows that the cross-attention method effectively utilizes the stability of the scattering center features.

[0115] Figure 8 is a diagram showing the analysis results of an ablation experiment of an overfitting strategy according to an exemplary embodiment. Figure 8 As shown in the figure, this part of the experiment analyzes the impact of overfitting on feature clustering by measuring the distance variance of the output features. Two characteristics of and The distance between them is:

[0116]

[0117] where ||·|| Euclidean The representative is the Euclidean distance. Since the fluctuation degree of feature clustering is more concerned in the incremental learning process, the average standard deviation of the features in each training period is expressed as follows:

[0118]

[0119] Among them, N CLASS represents the number of categories, and Representatives are instances The number of. Through the above formula, we can get Figure 8 , which shows the difference between having and not having dropout layers at different stages of training.

[0120] based on Figure 8 From the two accuracy change trends in , it can be seen that the network without the dropout layer fits faster than the network with the dropout layer. Then from the two standard deviation trends, the standard deviation of the network without the dropout layer is always lower than the standard deviation of the network with the dropout layer. This result verifies that removing the dropout layer structure helps to improve the stability of feature clustering. In addition, as the accuracy increases rapidly, the standard deviation also shows a rapid increase. Subsequently, the standard deviation gradually decreases as the training progresses, which shows that the late stopping strategy also helps to improve the stability of feature clustering.

[0121] Fig. 9is an ablation experiment analysis result diagram of another overfitting strategy according to an exemplary embodiment. The network training results with and without dropout layers are shown in FIG. Fig. 9 shown. Fig. 9 Conclusion and Figure 8 Similarly, removing the dropout layer and the late stopping strategy has a more obvious effect on feature clustering. This experimental result verifies the effects of our two methods from the perspective of clustering 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] Fig.10 is a diagram showing the analysis results of an ablation experiment of another overfitting strategy according to an exemplary embodiment. Fig.10 It can be seen that removing the dropout layer and the late stopping strategy both improve the performance of the incremental learning stage to varying degrees, among which only the late stopping strategy has the smallest improvement. It is worth noting that when using the initial network model, the highest classification accuracy in the basic stage reached 99.147%. The blue line with the best performance in the application of incremental learning is 98.540% in the basic stage. In other words, the overfitting method sacrifices the recognition performance to a certain extent in the basic stage in exchange for the improvement of incremental learning performance.

[0126] From the above four methods, it can be seen that the simultaneous removal of the dropout layer and the late stopping strategy only brings the smallest improvement, with an accuracy of 88.955% and 84.920% in the first and second phases, respectively, while the initial method has an accuracy of 88.578% and 84.662% in the first and second phases, respectively. However, the improvement brought by removing only the dropout layer is very obvious, with an accuracy of 93.099% and 87.827% in the first and second phases, respectively. This phenomenon is because the dropout layer slows down the network fitting process by possibly inactivating neurons, which is more effective than the effect of the late stopping strategy. However, comparing the gap between the blue and yellow lines with the gap between the green and red lines, it can be seen that the performance improvement brought by the late stopping strategy is different under different conditions. This is due to the randomness that the dropout layer brings to the training process, which greatly affects the learning process of the network.

[0127] For the angle constraint loss function proposed in this paper, the following experiments are conducted to verify the effectiveness of the angle constraint loss function in restricting the angle during network training and the improvement of incremental learning performance. This part mainly compares two methods, including using only the cross entropy loss function and using both the cross entropy loss function and the angle constraint loss function. For the second case, λ is set to 0.5, and the training process of the two methods uses the same hyperparameters, such as learning rate and optimizer. Considering that the value of the loss is difficult to control during the learning process, when the network training reaches different accuracy rates, we also compare the cosine value of the angle between the new class feature and the old classifier weight, which is called 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 should be noted that since the progress of network training is difficult to intervene manually, the test phase after each stage of training will first verify whether the accuracy exceeds the sampling accuracy to obtain the corresponding experimental results. However, for a sampling accuracy of 0.7, the accuracy of the two networks is very different, 0.704 and 0.749 respectively. This inevitably leads to a large error in the experimental results at a sampling accuracy of 0.7. It can be concluded from Table 3 that when the network trained with the angle constraint loss function reaches a certain accuracy, the angle between the new feature vector and the old classifier weight is smaller than that of the network without the angle constraint loss function, indicating that it is closer to orthogonality. This proves that during the network optimization process, our angle constraint loss function has a certain constraint on the angle.

[0131] Fig.11 is a diagram showing the analysis results of an ablation experiment of an angle constraint loss function according to an exemplary embodiment, see Fig.11 ,This section of the experiment 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 has slightly better results than the network without the angle constraint loss function. The biggest difference is in the fourth stage, where the two accuracies are 57.019% and 65.151%, respectively, with a difference of 8.13%. This part of the experiment proves that the introduction of the angle constraint loss function can improve the performance of incremental learning.

[0132] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to 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 further describe various possible combinations.

[0134] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. An incremental learning method, characterized in that: The method comprises: Extracting CNN features of the 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; Using the pre-built SCF-NET to fuse the CNN features and the electromagnetic scattering center features to obtain fused features; Regularizing the fused features, and constructing a loss function according to the regularized fused features; When performing incremental learning on the pre-constructed SCF-NET, the pre-constructed SCF-NET after adding the new category is optimized using the loss function, and the normalized weight of the new category is obtained according to the new category and the original category of the SCF-NET, and the pre-constructed SCF-NET is subjected to the next incremental learning according to the normalized weight of the new category, until the pre-constructed SCF-NET reaches a preset convergence condition, thereby obtaining the SCF-NET after incremental learning.

2. The incremental learning method according to claim 1, characterized in that: The CNN features for acquiring the SAR amplitude image and the estimated parameter set of the echo data corresponding to the SAR amplitude image include: Extracting CNN features of the SAR amplitude image through a simplified VGG network; The estimated parameter set corresponding to each electromagnetic scattering center in the SAR amplitude image is obtained through the 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, represents the estimated parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image, Indicates Φ when the formula in brackets reaches its minimum value i , Φ i represents the parameter set corresponding to the i-th electromagnetic scattering center in the SAR amplitude image, Represents the vector information obtained by converting the echo raw data corresponding to the SAR amplitude image, A i represents the amplitude value of the i-th electromagnetic scattering center, 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 feature is expressed as follows: in, represents the fusion feature, represents the fully connected layer in the SCF-NET, represents the normalized similarity matrix, Represents the electromagnetic scattering center feature.

5. The incremental learning method according to claim 4, characterized in that: The normalized similarity matrix is ​​expressed as follows: Among them, softmax(·) represents the softmax function, M s represents the similarity matrix, d k Representing the CNN features of the SAR amplitude image and the electromagnetic scattering center features Dimension.

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

7. The incremental learning method according to claim 6, characterized in that: The angle constraint loss function is expressed as follows: in, represents the kth row weight vector in the weight matrix of the classifier in SCF-NET, represents the fusion feature, and ||||2 represents 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 represents the probability that target i' is correctly classified, y i'c Represents the symbolic function corresponding to target i'.

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

Images