Radar target category incremental learning method based on multi-modal fusion

Abstract

The invention particularly relates to a radar target category incremental learning method based on multi-modal fusion. The method comprises the following steps: constructing a double-branch residual multi-scale feature extraction network; performing interaction and weighted fusion on the double-branch features through a cross-modal attention mechanism to obtain each modal enhanced feature; fusing the bimodal classification evidences based on a D-S evidence theory to obtain a fused classification result; in the basic category training stage, all basic category samples are used for training the feature extraction network and the fusion module; in a new category incremental learning stage, trained model parameters are frozen, trainable feature extraction branches are expanded for a new category, and joint training is carried out based on an example set and new category data; the current model is compressed into a lightweight model through knowledge distillation, and efficient updating of model parameters and knowledge migration are achieved. According to the method, the problems of disastrous forgetting, difficulty in multi-modal fusion modeling and low model updating efficiency in incremental learning are solved, and the target recognition accuracy and the model adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, and more specifically to an incremental learning method for radar target categories based on multimodal fusion. Background Technology

[0002] Radar (Radio Detecting and Ranging) is a widely used electronic detection device that uses electromagnetic waves to detect targets and acquire relevant information. The transmitter of a radar system generates high-frequency electromagnetic waves, which are emitted into space through the antenna. Upon contact with a target object, the waves are reflected, and the intensity of the reflected echo is affected by factors such as the material and geometry of the target object and the radar's observation angle. By capturing these echoes, the receiver performs a series of signal processing steps, including signal amplification, time-frequency analysis, and noise suppression, to deduce important information such as the target's coordinates, speed, shape, and size in space. Therefore, radar plays a crucial role. The rapid development of information intelligence has led to Radar Automatic Target Recognition (RATR) technology gradually becoming a research hotspot in the field of radar pattern recognition. The tasks of radar automatic target recognition can be roughly divided into coarse-grained to fine-grained tasks: (1) Target Discrimination; (2) Target Classification; and (3) Target Identification. The three levels mentioned above correspond to: sensing the attributes of the target, classifying the target into categories, and identifying the specific model of the target. With the increasing demand for finer-grained identification, more and more advanced technologies are being incorporated into radar automatic target identification.

[0003] Based on the pulse width of the transmitted radar signal, signals can be broadly categorized into two types: narrowband signals and wideband signals. A typical example of a narrowband signal is Jet Engine Modulation (JEM), often used to capture the micro-motion characteristics of rotating components of a target. Wideband signals include High Resolution Range Profiles (HRRP), Synthetic Aperture Radar (SAR) images, and Inverse Synthetic Aperture Radar (ISAR) images. HRRP signals are the vector sum of the one-dimensional projections of the target's scattered point echoes onto the radar's line-of-sight, containing rich geometric information about the target. Compared to the complex imaging principles of SAR and ISAR, the processing of HRRP signals is relatively simple, leading to their widespread use in automatic target identification. HRRP and JEM signals emphasize different aspects of the same target representation, providing complementary information and effectively compensating for the lack of information in single-mode target description.

[0004] Most existing deep learning-based radar target recognition training methods employ a static approach, involving closed-loop training on a fixed database to optimize model performance. This method relies on a large amount of pre-collected data that remains constant throughout the training process. However, in real-world scenarios, data on the same target exhibits multimodal characteristics, and this data is continuously collected over time, creating a multimodal category incremental task. Models trained using traditional methods have fixed parameters and cannot simultaneously integrate multimodal data and continuously learn like the human brain. Therefore, when facing multimodal category incremental tasks, training the model solely on new tasks leads to a sharp decline in performance on older tasks, resulting in catastrophic forgetting. Effective integration of different modalities is also necessary. Existing methods typically require updating the database and directly fusing multimodal information to maintain high model performance before retraining the model. However, this undoubtedly incurs higher time, storage, and computational costs, making it difficult to meet the demands of rapid response in real-time scenarios. How to enable continuous learning of models in multimodal class incremental tasks is also a problem that Class-Incremental Learning (CIL) needs to solve.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides an incremental learning method for radar target categories based on multimodal fusion, which addresses the shortcomings of existing incremental learning methods that focus on single-modal data, making it difficult to utilize the complementary information of multimodal radar signals. When dealing with tasks where radar target categories are continuously added, these methods suffer from catastrophic forgetting, difficulties in multimodal fusion modeling, and low model update efficiency, resulting in limited target recognition performance.

[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to a first aspect of the present invention, a radar target category incremental learning method based on multimodal fusion is provided, the method comprising: A dual-branch residual multi-scale feature extraction network is constructed to extract features from JEM and HRRP signals respectively. JEM signals are used to describe the micro-motion characteristics of the rotating parts of the target, while HRRP signals are used to reflect the scattering intensity distribution of the target scatterer along the radar line of sight. By using a cross-modal attention mechanism to interact with and weight the bi-branch features, enhanced features for each modality are obtained. Based on the DS evidence theory, the classification evidence of the two modes is fused to obtain the fused classification result; During the basic category training phase, the feature extraction network and fusion module are trained using all basic category samples; In the incremental learning phase for new categories, the parameters of the trained model are frozen, the trainable feature extraction branch is expanded for new categories, and the model is jointly trained based on the example set and the new category data. Knowledge distillation compresses the current model into a lightweight model, enabling efficient updating of model parameters and knowledge transfer.

[0009] In some exemplary embodiments, in the dual-branch residual multi-scale feature extraction network, the JEM branch contains 5 one-dimensional residual blocks, the HRRP branch contains 4 one-dimensional residual blocks, and each residual block is followed by a max pooling layer and a channel attention module.

[0010] In some exemplary embodiments, the cross-modal attention mechanism achieves the interaction and enhancement of features between modalities by multiplying the channel attention weights of one modality with the feature map of another modality.

[0011] In some exemplary embodiments, the fusion method based on DS evidence theory includes: The classification outputs of each modality are modeled as a Dirichlet distribution; Calculate the classification confidence and uncertainty for each mode; The confidence and uncertainty of the two modes are fused using the DS fusion rule to obtain fused classification evidence.

[0012] In some exemplary implementations, the example set is constructed by selecting samples that are closest to the feature centers of each category, in order to retain old category knowledge during the incremental phase.

[0013] In some exemplary embodiments, the incremental learning phase employs a dynamic network expansion mechanism, specifically including: Freeze the parameters of the old model; Add trainable feature extraction branches and classifiers for new categories; We use joint training with data from both the old and new classes, and introduce distillation loss to preserve the performance of the old class.

[0014] In some exemplary embodiments, the distillation loss is calculated based on the difference in output distribution between the old and new models in each modal branch, and is used to constrain the new model to retain knowledge of the old categories.

[0015] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the radar target category incremental learning method based on multimodal fusion described in the first aspect.

[0016] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the radar target category incremental learning method based on multimodal fusion described in the first aspect above.

[0017] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the radar target category incremental learning method based on multimodal fusion as described in the first aspect by executing the executable instructions.

[0018] The incremental learning method for radar target categories based on multimodal fusion provided in the embodiments of the present invention solves the problems of catastrophic forgetting, difficulty in multimodal fusion modeling, and low model update efficiency in incremental learning. It fully leverages the complementary value of multimodal data, enabling continuous expansion and recognition of radar target categories, improving recognition accuracy and model adaptability, and is suitable for scenarios with high real-time requirements. Compared with existing technologies, it has the following beneficial effects: 1. Enhanced multimodal feature extraction capability: By designing a dual-branch residual multi-scale feature extraction network and introducing a cross-modal decision fusion mechanism, the complementary information of different modal radar signals (such as JEM and HRRP) can be fully explored and fused, significantly improving the model's ability to express target features and its recognition accuracy.

[0019] 2. Application of Multimodal Radar Signals in Incremental Learning: By combining information from different modes, multimodal radar signals can effectively compensate for the shortcomings of a single signal source. In category incremental learning tasks, this improves model recognition performance while combating catastrophic forgetting.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 Flowchart of the incremental learning method for multimodal radar categories; Figure 2 This is a dual-branch radar signal feature extraction network; Figure 3 A multimodal fusion diagram of the DS evidence theory; Figure 4 Schematic diagram of model feature enhancement principle; Figure 5 This is a schematic diagram of the knowledge distillation principle for the model. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0024] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0025] In modern applications, radar plays an increasingly important role. Deep learning-based methods, through deep neural networks, can fully mine the potential patterns in data to obtain more accurate radar target representation features, thereby achieving high-precision identification tasks. In practical applications, data is continuously collected over time, and the database is constantly updated. When a new target category appears, the model needs to be retrained. However, if only samples of the new category are used for training, the model will suffer from catastrophic forgetting. Incremental learning methods can effectively overcome catastrophic forgetting. At the same time, most existing incremental learning methods focus on processing single-modal data, neglecting the modal diversity of radar signal data. For example, JEM signals and HRRP signals of the same target category reflect different focuses of the target, possessing independence and complementarity. Therefore, efficiently utilizing multimodal radar signals to improve target recognition accuracy during incremental learning has become an urgent problem to be solved in the field of automatic radar target recognition.

[0026] Most existing incremental learning methods focus on processing single-modal data, neglecting the modal diversity of radar signal data. These modalities reflect different aspects of the target and possess both independence and complementarity. Therefore, efficiently utilizing multimodal radar signals to improve target recognition accuracy during incremental learning has become an urgent problem to be solved in the field of automatic radar target recognition.

[0027] (1) The problem of catastrophic forgetting When a radar recognition model introduces new categories for incremental learning, the model is prone to forgetting the previously learned categories, leading to a decline in recognition performance.

[0028] (2) There is a lack of research on the application of incremental learning in multimodal radar data. Most current incremental learning methods are mainly for single-mode data, making it difficult to fully utilize the complementary information between multiple modes (such as JEM and HRRP) in radar signals.

[0029] (3) Modeling and fusion of differences and complementarities between radar signal modes Different modes of radar data reflect different physical characteristics of the target. How to effectively fuse them without sacrificing the independence of their respective characteristics remains a challenge.

[0030] (5) The problem of efficient model updating under continuous data growth In practical applications, the radar database is constantly expanding, requiring the design of model structures and training strategies that support dynamic expansion and rapid adaptation to new categories, in order to avoid frequent retraining of the overall model.

[0031] To address the shortcomings and deficiencies of existing technologies, this example implementation provides an incremental learning method for radar target categories based on multimodal fusion, which has the following characteristics: First, a multimodal radar signal fusion mechanism: A dual-branch residual multi-scale network is used to process different modal radar signals (such as JEM and HRRP). A cross-modal decision fusion module is introduced to achieve effective fusion of high-level semantic information between modalities. Second, a multimodal dynamic network structure expansion mechanism: In the incremental stage, the original model parameters are frozen, and trainable branches are introduced only for new categories, achieving scalability of the model structure and isolated learning between old and new tasks. Combined with multimodal radar signals, this improves model recognition performance while combating catastrophic forgetting. Specifically: During the basic category model training phase, a recognition network structure based on multimodal information fusion is adopted. Radar signal data of different modalities (including but not limited to JEM signals and HRRP signals) are input into a two-branch residual multi-scale feature extraction network for processing to extract multimodal features. The extraction results are then fused through a cross-modal decision fusion mechanism.

[0032] During training, multiple loss functions are introduced to jointly optimize the feature extractor and classifier, thereby improving the model's ability to discriminate target category features. After each training round, based on the distance metric between sample features and the center of each category, the samples closest to the feature center of each category are selected to form an exemplar set. This exemplar set is used to retain old category knowledge in subsequent incremental phases.

[0033] In the incremental learning phase for new categories, the model is updated by combining a dynamic network expansion mechanism and a knowledge distillation strategy. Specifically, this includes freezing the parameters of the already trained old category model to maintain its ability to recognize old categories, while simultaneously extending a new trainable model branch for the newly added categories. This new model is trained using both the new category data and the example set to achieve a unified recognition capability for both old and new categories.

[0034] After training, a knowledge distillation process is introduced. The currently trained model is used as the teacher model, and distillation guides the student model to perform parameter compression and knowledge transfer, thereby obtaining a compact model that balances the recognition performance of both new and old categories. This process can be executed cyclically in application scenarios where new categories are continuously added, enabling the gradual expansion of radar target category recognition.

[0035] The steps in this exemplary embodiment will now be described in more detail with reference to the accompanying drawings and embodiments.

[0036] Step 1: Radar target feature extraction based on residual multi-scale A residual multi-scale radar target feature extraction method is proposed. This method integrates residual networks and attention mechanisms, fully combining the powerful feature learning capabilities of residual networks and the attention mechanism's ability to focus on important information, to effectively extract global and local information from target data. This design not only reduces the loss of target information but also improves network performance, providing more robust feature representations for radar target identification. Specifically, the model utilizes signal data from both JEM and HRRP modalities simultaneously by combining one-dimensional convolution and channel compression attention mechanisms. Independent feature extraction branches are designed for each modality to ensure more accurate feature extraction. Furthermore, to enable complementary information between the two modalities to work synergistically and reduce the impact of feature conflicts between modalities, cross-modal weights are introduced between the two branches. This aims to integrate the most valuable features for the recognition task and weaken conflicting features. The JEM branch includes five one-dimensional residual blocks, while the HRRP branch includes four one-dimensional residual blocks, each followed by a max-pooling layer. The nth... i The output of each residual block for: (1) in No. i The input of each residual block, Represents a convolutional network. express Convolution is used to adjust the number of channels. Each max-pooling layer is followed by a channel attention layer to capture the weights of channels in each modality, thereby enhancing the focus on the target attribute. The calculation formula is as follows: (2) in This represents the feature map of the c-th channel. This represents the global feature value of the c-th channel. This is the weight matrix of the first fully connected layer. This is the weight matrix of the second fully connected layer. It is the ReLU activation function. It is the Sigmoid activation function. Simultaneously, the weights of one modality are fused with features from another modality to achieve mutual attention within the network, thereby integrating features and mitigating conflicts. Then the... Input of each residual block for: (3) in Indicates the max pooling layer. These are the weights generated by SENet for the corresponding layer of the JEM branch. These are the weights generated by SENet for the corresponding layer of the HRRP branch. Since the JEM branch has one more residual block than the HRRP branch, the output of the last residual block of the JEM branch uses the weights generated by the last SENet from the HRRP branch. Finally, each branch follows a fully connected layer, which unfolds the feature map into a one-dimensional vector for subsequent processing.

[0037] Step 2: Multimodal fusion strategy based on DS evidence theory Radar signals encompass various types, which are physically independent but can provide complementary characterization information for the same target. This invention selects two radar signals with typical characteristics: the JEM signal, which describes the micro-motion characteristics of the target's rotating components; and the HRRP signal, which reflects the scattering intensity distribution of the target scatterer along the radar line of sight. These two signals are independent yet complementary, providing an ideal data foundation for fusion. Category incremental learning, as a technique simulating the human brain's learning process, should be able to effectively integrate multimodal information when faced with multimodal data in the real world. Based on this idea, and combined with a multimodal fusion strategy based on DS evidence theory, the JEM and HRRP modalities are effectively complementary.

[0038] Specifically, the JEM and HRRP signals are each processed by a residual multi-scale feature extraction network, and then output by a classifier to obtain classification evidence for the two corresponding modalities, i.e., the measures supporting classification, denoted as [reference 1]. and Where M represents the number of currently known categories, and the superscripts represent the JEM and HRRP modalities, respectively. Dirichlet distribution modeling is performed, evidence... and As the parameter corresponding to the Dirichlet distribution.

[0039] (4) Finally, the Dirichlet distributions corresponding to the two modes were obtained, respectively and Finally, the classification confidence scores for each category are obtained using the following formula. and uncertainty The relationship between the two satisfies formula (\ref{3-6}), which maps the Dirichlet distributions of the two modes to respectively: and .

[0040] (5) (6) Then, the classification confidence of the bimodal model is... and uncertainty Fuse according to the formula: (7) Where C is the conflict metric between the two modes. This is used to standardize the classification probabilities and uncertainties after fusion. Finally, the Dirichlet distribution corresponding to the fused mode and the classification evidence for the fused mode are derived. This is used to determine the final classification result.

[0041] Step 3: Incremental learning of radar targets based on multimodal fusion Dynamic network-based methods effectively retain old knowledge and overcome catastrophic forgetting by dynamically extending model branches for new categories in incremental tasks and freezing parameters learned from previous tasks. Although this approach increases computational costs, in real-world radar target recognition scenarios, dynamic networks outperform playback-based and regularization-based incremental learning methods as the number of tasks increases and high recognition performance is required. However, most current category incremental learning methods focus on a single modality and fail to adequately consider the needs of multimodal data processing in the real world. To address this issue, a multimodal fusion-based radar target category incremental learning method is proposed. This method combines a dynamic network-based category incremental learning approach with a multimodal fusion strategy based on DS evidence theory, fully leveraging the ability of dynamic networks to maintain high performance as tasks increase while effectively integrating the complementary characteristics of multimodal data. Through this design, the model can better handle incremental tasks, adapt to more realistic environments, and exhibit excellent performance stability. The main process of this method consists of two parts: model training for the basic categories and incremental learning for new categories.

[0042] During training with the base categories, half of all categories are divided into base categories. JEM and HRRP signals are fed into the residual multi-scale feature extraction backbone network, followed by multimodal fusion based on DS evidence theory to obtain the final target category prediction result. Similar to traditional recognition tasks, the loss is calculated using the cross-entropy loss function, and the parameters are optimized through backpropagation.

[0043] Compared to other incremental learning methods, dynamic networks strike a better balance between model stability and adaptability. By resetting the feature extractor for the new category at each incremental task stage, the model gains the ability to learn new categories while freezing the feature extractors for the old categories from the previous stage to preserve knowledge of those categories. This addresses the challenge of maintaining high recognition performance despite increasing task demands. Inspired by FOSTER, the incremental stage is also a continuous improvement process. For example, during training on older tasks, facial features of different animals can serve as key differentiators, and the model ignores common stripe features between categories. When a new task arrives, stripe features become the key differentiator again. Therefore, the improvement process is how the new model better integrates key features from all known categories. The following section will detail FOSTER and its improvements.

[0044] First, decompose the model F to obtain... ,in For feature extractor, Let N be the classifier, and N be the number of classes. In the t-th increment stage, the old model from the previous stage... It will be frozen, thus ensuring that the learned knowledge is not forgotten. , Indicates the quantity of the old category, and For ease of subsequent representation, the following settings are made: (1) When the input is unimodal, the classifier only includes the branch corresponding to that modality; (2) When the input is multimodal, the classifier includes branches of all modalities. Meanwhile, the new model... We need to fit the residual between the old model output and the old model output. It can be represented as: (8) Among them, the newly added classifier Decomposed into and , The number of all known categories For the newly added feature extractor. The parameters are Meanwhile, the entire optimization process of the model can be represented as: (9) In this structure, to ensure the recognition accuracy of all categories, cross-entropy loss is used to calculate the current model. The loss is calculated by the following formula, which represents the difference between the output after multimodal fusion and all category labels: (10) In addition, the new feature extractor Not only does it need to learn new categories, but it also cannot neglect learning from old categories. Without the constraint of old categories, the model might only focus on new categories, leading to a decline in recognition performance on old categories. Therefore, cross-entropy loss is used to guide and force the new feature extractor to learn new categories. Distinguish among all known categories and enhance the representation ability of each branch for all known categories. The loss of each branch is as follows: (11)

[0045] (12) Meanwhile, the feature representations of the old categories cannot be learned well in the process of learning the new categories. Therefore, for the new model... For the output distribution of the old categories (specifically including the JEM and HRRP branches), a distillation loss is introduced to make it as similar as possible to the corresponding old model. The output distributions are similar. The distillation losses for each branch are as follows: (13) (14) in and They represent the new model respectively The output of the new feature extractor after passing through the JEM and HRRP branches can be used to obtain the total loss in the incremental stage as follows: (15) After the incremental phase ends, the current model As a teacher model, it utilizes the outputs of its two branches and the output after decision fusion, through loss... , and Guiding student models The total loss during the model compression phase is as follows: (16) The method of the present invention is further illustrated by simulation experiments, as shown in Table 1.

[0046] Table 14 Comparative Experiment Results of Incremental Stages

[0047] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments.

[0048] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0049] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0050] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0051] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A radar target category incremental learning method based on multimodal fusion, characterized in that, The method includes: A dual-branch residual multi-scale feature extraction network is constructed to extract features from JEM and HRRP signals respectively. JEM signals are used to describe the micro-motion characteristics of the rotating parts of the target, while HRRP signals are used to reflect the scattering intensity distribution of the target scatterer along the radar line of sight. By using a cross-modal attention mechanism to interact with and weight the bi-branch features, enhanced features for each modality are obtained. Based on the DS evidence theory, the classification evidence of the two modes is fused to obtain the fused classification result; During the basic category training phase, the feature extraction network and fusion module are trained using all basic category samples; In the incremental learning phase for new categories, the parameters of the trained model are frozen, the trainable feature extraction branch is expanded for new categories, and the model is jointly trained based on the example set and the new category data. Knowledge distillation compresses the current model into a lightweight model, enabling efficient updating of model parameters and knowledge transfer.

2. The method according to claim 1, characterized in that, In the dual-branch residual multi-scale feature extraction network, the JEM branch contains 5 one-dimensional residual blocks, and the HRRP branch contains 4 one-dimensional residual blocks. Each residual block is followed by a max pooling layer and a channel attention module.

3. The method according to claim 1, characterized in that, The cross-modal attention mechanism achieves feature interaction and enhancement between modalities by multiplying the channel attention weights of one modality with the feature map of another modality.

4. The method according to claim 1, characterized in that, The fusion method based on DS evidence theory includes: The classification outputs of each modality are modeled as a Dirichlet distribution; Calculate the classification confidence and uncertainty for each mode; The confidence and uncertainty of the two modes are fused using the DS fusion rule to obtain fused classification evidence.

5. The method according to claim 1, characterized in that, The example set is constructed by selecting samples that are closest to the feature centers of each category, and is used to retain old category knowledge during the incremental phase.

6. The method according to claim 1, characterized in that, The incremental learning phase employs a dynamic network expansion mechanism, specifically including: Freeze the parameters of the old model; Add trainable feature extraction branches and classifiers for new categories; We use joint training with data from both the old and new classes, and introduce distillation loss to preserve the performance of the old class.

7. The method according to claim 1, characterized in that, The distillation loss is calculated based on the difference in output distribution between the old and new models in each modal branch, and is used to constrain the new model's retention of knowledge of the old categories.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the radar target category incremental learning method based on multimodal fusion as described in any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the radar target category incremental learning method based on multimodal fusion as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the radar target category incremental learning method based on multimodal fusion as described in any one of claims 1 to 7 by executing the executable instructions.

Images